Scoring
How Each Mechanism Was Scored
The criteria every mechanism is scored against, then the findings behind each score: what the mechanism is, where it lives in the brain, how it has been measured in people, and how it has been measured in AI.
Scoring Criteria
Level of Measurement: Humans
Every mechanism here is established in humans, so the question is only how directly.
- Directly Measured
- Psychologists have a standard experiment built to test this ability specifically, and its results have been replicated or pooled in a meta-analysis.
- Indirectly Measured
- The ability is known from other kinds of evidence, such as diary studies, people with brain injuries, or experiments designed to test something else, rather than from a test built for it.
Level of Measurement: AI
An AI system can hold information in three places: what the model learned in training, what it can see in the current conversation, and any saved memory it can look things up in. Each was checked.
- Directly Measured
- A published study tests AI on this ability by name, or gives AI the same experiment psychologists use to test it in people.
- Indirectly Measured
- Studies have tested something close to this ability, but under a different name, as part of a broader test, or in only some of the places an AI system holds information.
AI Memory Performance
Rating AI against people takes results from both on the same ability. Those results can be compared in two ways.
- Direct comparison
- People and AI took the same test, with the same questions, scored the same way, and neither side could look back at the material unless the other could too.
- Approximate comparison
- People and AI took different tests of the same ability, but the results can be put on a common scale, such as accuracy above chance or a standard effect size. A rating is given only when the result is clear despite the differences between the tests, and it is labelled approximate.
Either way, two conditions apply:
- 1 The human results are documented. The study reports how many people were tested and how.
- 2 An uneven test only counts against its advantage. When a test was easier for one side, a result that goes against that advantage counts, and so does a tie. An AI that could still see the material and scored below people is rated Below Human. One that could see it and only matched people is rated Comparable, and always marked provisional. One that could see it and scored above people is not counted.
The ratings:
- Exceeds Human
- AI did better than the normal range for people.
- Comparable to Humans
- AI fell within the normal range for people.
- Below Human
- AI did worse than the normal range for people.
- Not Yet Compared
- Both people and AI have been tested on this ability, and the two could be put on one scale, but no study has yet done so with documented human results. An AI benchmark that prints a human score without saying who was tested or how does not count as that study.
- Comparison Not Possible
- No fair test could put people and AI on one scale for this ability: what it measures in people has no counterpart that could be tested the same way in AI.
There is no Absent rating. When nothing in an AI system does what an ability does, there is nothing to compare, and the rating is Comparison Not Possible.
- The normal range for people
- The spread the human studies report: a standard deviation, a confidence interval, or the range across groups. If a study gives only an average, a difference counts only when the researchers' own statistics find one.
- Which AI system
- When a study tests several AI systems, the best one is rated, since the question is what today's AI memory can do. How much the systems differed is noted alongside.
- Provisional
- A rating is marked provisional when it rests on one study, on studies not yet peer reviewed, on a single AI system, or on a tie from a test that was easier for the AI.
- When results disagree
-
- 1 Any comparison outranks none. A result that can be compared, directly or approximately, takes precedence over Not Yet Compared and Comparison Not Possible, which record that no comparison was made, not that one came out a particular way.
- 2 Where comparisons disagree, a direct comparison outranks an approximate one. Between comparisons of the same strength, the one whose test better captures the ability as psychologists study it in people stands. The others are still shown.
- 3 Where no test is clearly the better one and the comparisons point in different directions, the more cautious rating stands, meaning the lower one, and it is marked provisional.
Definitions, what each score requires, and worked examples are on the Method page.
Table 1a, Scored Against These Criteria
Where each mechanism lands when scored by the criteria above. Select a dot to jump to that mechanism's findings. Hollow dots are being re-checked.
Measurement in AI
Not possible Comparison Not Possible (6)
Not yet Not Yet Compared (9)
Below Below Human (8)
Comparable Comparable to Humans (5)
Exceeds Exceeds Human (1)
not on the scale
AI Memory Performance
- Directly
- Tested in AI by name, or with the human experiment.
- Indirectly
- Tested under another name, or only in part.
- Comparison Not Possible
- No fair test could put both on one scale.
- Not Yet Compared
- Could be compared; no study has done it yet.
- Below Human
- AI did worse than people.
- Comparable to Humans
- AI fell within the range for people.
- Exceeds Human
- AI did better than people.
Level of measurement
AI Memory Performance
- Memory Systems
- Representational Structures
- Memory Processes
- Metacognition
Each dot is one of the 29 mechanisms. Where a dot sits within a cell carries no meaning, just which cell it is in.
What Each Relationship Needs, Against How AI Measures Up
Every mechanism in one bar, coloured by how AI memory compares with people. Under it, a filled cell marks each mechanism an advice and consulting relationship or a therapy relationship draws on (Table 2). Select a block or a mark to jump to that mechanism.
Better or Same as Humans
Worse than Human Memory
Not Compared
Semantic Knowledge Confidence Calibration Instrumental Learning Working Memory Priming Retrieval Interference Partner Model Episodic Memory Standing Constraints Temporal Placement Metacognitive Control Schema Formation Classical Conditioning Procedural Memory Reality Monitoring Source Monitoring Schema Application Integrative Inference Prospective Initiation Recognition Feeling of Knowing Consolidation Context-Dependent Retrieval Autobiographical Memory Involuntary Retrieval Reconsolidation Non-Associative Learning Forgetting Motivated ForgettingAI Memory Performance
- Exceeds Human
- Comparable to Humans
- Below Human
- Not Yet Compared
- Comparison Not Possible
- Advice and Consulting draws on 13 of 29 mechanisms: 4 at or above people, 3 below people, 5 not yet compared, and 1 can't be compared.
- Therapy draws on 23 of 29 mechanisms: 5 at or above people, 6 below people, 9 not yet compared, and 3 can't be compared.
Show as a table
| Mechanism | AI memory performance | Advice and Consulting (elements of 9) | Therapy (elements of 15) |
|---|---|---|---|
| Semantic Knowledge | Exceeds Human | 3 | 1 |
| Confidence Calibration | Comparable to Humans | 3 | 2 |
| Instrumental Learning | Comparable to Humans | 1 | 3 |
| Working Memory | Comparable to Humans | 1 | 2 |
| Priming | Comparable to Humans | – | 1 |
| Retrieval Interference | Comparable to Humans | – | – |
| Partner Model | Below Human | 6 | 13 |
| Episodic Memory | Below Human | 4 | 13 |
| Standing Constraints | Below Human | 4 | 6 |
| Temporal Placement | Below Human | – | 8 |
| Metacognitive Control | Below Human | – | 3 |
| Schema Formation | Below Human | – | 2 |
| Classical Conditioning | Below Human | – | – |
| Procedural Memory | Below Human | – | – |
| Reality Monitoring | Not Yet Compared | 2 | 5 |
| Source Monitoring | Not Yet Compared | 3 | 3 |
| Schema Application | Not Yet Compared | 2 | 2 |
| Integrative Inference | Not Yet Compared | 2 | 2 |
| Prospective Initiation | Not Yet Compared | – | 3 |
| Recognition | Not Yet Compared | – | 2 |
| Feeling of Knowing | Not Yet Compared | 1 | 1 |
| Consolidation | Not Yet Compared | – | 1 |
| Context-Dependent Retrieval | Not Yet Compared | – | 1 |
| Autobiographical Memory | Comparison Not Possible | 3 | 4 |
| Involuntary Retrieval | Comparison Not Possible | – | 4 |
| Reconsolidation | Comparison Not Possible | – | 1 |
| Non-Associative Learning | Comparison Not Possible | – | – |
| Forgetting | Comparison Not Possible | – | – |
| Motivated Forgetting | Comparison Not Possible | – | – |
Memory Systems · Immediate
Working Memory
Definition
- Working memory is a limited-capacity system that holds information briefly and works on it (updating, reordering, combining), described in the multicomponent model as a set of storage subsystems coordinated by attentional control (Baddeley, 2012).
- Its storage is tightly capped: pooled evidence across paradigms points to a central limit of about four chunks when rehearsal and grouping are prevented (Cowan, 2001).
- Working memory is separable from simple short-term storage: in 133 adults, working memory and short-term memory formed distinct but related factors, and only working memory predicted fluid intelligence, which the authors attributed to keeping information active despite interference (Engle et al., 1999).
- It differs from long-term memory in being a temporary workspace, though an added episodic buffer component is proposed to bind information from the subsystems with knowledge retrieved from long-term memory (Baddeley, 2000).
Neural Substrate
- In monkeys doing a delayed-response task, 87 of 288 recorded dorsolateral prefrontal neurons changed firing during 1–6 second delays, and 79% of those were tuned to the remembered location, the classic evidence for persistent delay activity (Funahashi et al., 1989).
- In a lesion study of 19 patients with dorsolateral prefrontal damage, 152 with other lesions and 54 controls, dorsolateral prefrontal damage impaired manipulating verbal and spatial information in working memory, with the left side necessary for manipulation (Barbey et al., 2013).
- In fMRI, activity in the posterior parietal cortex was tightly correlated with the limited amount of information people could hold in visual short-term memory (Todd & Marois, 2004).
- Human and primate evidence shows working memory contents stored across many regions, from sensory to parietal to prefrontal cortex, rather than in the prefrontal cortex alone (Christophel et al., 2017).
Human Measurement
Level of Measurement: Directly Measured
- In the change-detection paradigm, people could hold only about four colours or orientations at once, but could hold four objects with several features each, showing capacity is counted in integrated objects (Luck & Vogel, 1997).
- The two most common working memory paradigms, complex span tasks and the n-back task (say whether the current item matches the one n items back), were only weakly correlated across pooled studies, with significant heterogeneity, so they cannot be used interchangeably (Redick & Lindsey, 2013).
- A meta-analysis of 86 samples found an average true-score correlation of .479 between working memory and general intelligence, showing the measures are related but not the same construct (Ackerman et al., 2005).
- Across three experiments, n-back performance predicted fluid intelligence, especially at higher loads, but its reliability was too low for it to serve as a good measure of individual differences (Jaeggi et al., 2010).
- A meta-analysis of 24 neuroimaging studies (668 activation coordinates) confirms the n-back as a standard, replicated working memory paradigm, consistently engaging lateral prefrontal, premotor and posterior parietal cortex (Owen et al., 2005).
AI Measurement
Level of Measurement: Directly Measured
AI Memory Performance: Comparable to Humans approximate provisional
- Given the n-back task one letter or grid per turn, ChatGPT (GPT-3.5) fell to a sensitivity of about d′ = 1 at 3-back, the load at which human performance from a published n-back study also drops, which the authors called a capacity limit strikingly similar to humans; working memory's AI counterpart is the context window (what the model can see in the current conversation), since training and saved memory are long-term stores (Gong et al., 2024).
- The comparison is approximate and uneven: a chat model can answer n-back only because earlier letters stay in its context window, so it can look back while people must hold them in mind, and the human data come from separate timed lab studies; because GPT-3.5 still did no better than people, the result counts and sits within the human range, which is the basis for Comparable (Gong et al., 2024).
- Ten open-weight models (1–27 billion parameters) given a multi-turn n-back, with all earlier letters visible, mostly approached chance by 3- or 4-back and often fell below a human reference pooled from five studies, while the largest (Qwen 3.5 27B) exceeded humans; that above-human result is not counted, because the model could see the material people had to remember (Xiong et al., 2026).
- Today's strongest systems score far above people only on uneven tests: in a 2026 preprint, 100 online participants had median accuracy of 0.6 on letter serial recall while ten frontier models (e.g., GPT-5.2, Claude Opus 4.1, Gemini 3 Pro) scored near 1.00 with no serial-position or list-length effects, and a 2024 preprint put models at or above the 99.5th percentile on the WAIS-IV Working Memory Index, but in both the letters or digits stayed visible in the prompt (Schug & Lake, 2026).
- Spread across systems is wide: on n-back up to 10-back, the best of eight models (Llama 3.1 70B) kept 93% retrieval accuracy at 3-back and 57% at 10-back while smaller open models and GPT-3.5 fell in lower performance tiers, and the authors traced many failures to misunderstanding or losing track of the task rather than to storage limits (Hu & Lewis, 2025).
- When information must be held without being written into the conversation, models fail: across 17 models, most could not keep a secretly chosen number consistent across questions, and a hidden-arithmetic task was solved by at most 26% without step-by-step writing, rising to 100% only when a reasoning model wrote its steps out (a preprint, with no human data) (Huang et al., 2025).
- Adapting the human proactive-interference paradigm to 30+ models, a workshop paper found retrieval of the latest value of a repeatedly updated item declined log-linearly toward zero as earlier values accumulated; larger models resisted longer, while context-window length had no significant effect (no human data) (Wang & Sun, 2025).
Memory Systems · Declarative
Semantic Knowledge
Definition
- Tulving proposed semantic memory as one of three memory systems (procedural, semantic, episodic), with episodic memory a specialized subsystem of semantic memory, and argued that the systems differ in the kind of conscious awareness that goes with retrieval: knowing a fact versus re-experiencing the occasion it was learned (Tulving, 1985).
- Semantic cognition is the ability to use, manipulate and generalize knowledge acquired over the lifespan, and the controlled semantic cognition framework separates the stored conceptual knowledge itself from the control processes that retrieve and shape it for a given task (Lambon Ralph et al., 2017).
- A 2019 review concludes that semantic and episodic memory are intertwined and their neural correlates overlap considerably, yet the two still retain a measure of distinctiveness, so the 1972 distinction is kept but no longer treated as clear-cut (Renoult et al., 2019).
Neural Substrate
- A meta-analysis of 120 functional neuroimaging studies of semantic processing found a left-lateralized network of seven regions: posterior inferior parietal lobe, middle temporal gyrus, fusiform and parahippocampal gyri, dorsomedial and ventromedial prefrontal cortex, inferior frontal gyrus, and posterior cingulate gyrus (Binder et al., 2009).
- In five patients with focal temporal lobe atrophy, knowledge of word meanings and of living and man-made things progressively dissolved while day-to-day episodic memory, grammar, and perceptual abilities were relatively preserved, a syndrome the authors named semantic dementia (Hodges et al., 1992).
- Three people with bilateral hippocampal damage from birth or early childhood had severe amnesia for everyday events yet attended mainstream schools and reached low-average to average levels of language, literacy and factual knowledge, suggesting semantic learning is not fully hippocampus-dependent (Vargha-Khadem et al., 1997).
- A contrasting review of amnesia after medial temporal lobe or diencephalic damage found no compelling evidence that semantic memory is spared relative to episodic memory, arguing both forms of declarative memory depend on these structures (Squire & Zola, 1998).
Human Measurement
Level of Measurement: Directly Measured
- Category (semantic) fluency, naming as many members of a category as possible in a fixed time, is a standard semantic memory test; a meta-analysis of 153 studies with 15,990 participants found Alzheimer's patients more impaired on semantic than on letter fluency (r = .73 vs .57), consistent with degradation of the semantic store (Henry et al., 2004).
- Semantic priming, faster responses to a word after a meaning-related word, is a standard laboratory measure of semantic memory; a meta-analysis of 26 studies found reliable priming from purely semantic relations, though smaller on average than when the words are also associated (Lucas, 2000).
- Vocabulary tests give a replicated picture of semantic knowledge across adulthood: across 324 younger-older comparisons in 210 articles, older adults scored higher by an average of 0.80 SD (Verhaeghen, 2003).
- In a population-based longitudinal sample of 829 adults aged 35 to 80, semantic memory showed small gains until about age 55 and smaller late-life declines than episodic memory (Rönnlund et al., 2005).
- The Information subtest of the Wechsler Adult Intelligence Scale (WAIS-IV), which asks general-knowledge questions, is normed on a standardization sample of 2,200 adults, with subtest scaled scores set to a mean of 10 and a standard deviation of 3 (Block et al., 2025).
AI Measurement
Level of Measurement: Directly Measured
AI Memory Performance: Exceeds Human provisional
- A 2024 preprint gave 10 AI chat models the WAIS-IV Information subtest, general-knowledge questions such as "Who was the first president of the United States?", using the manual's instructions and items, with answers scored by clinical psychologists against norms for ages 25 to 29; 9 of the 10 models earned a scaled score of 19, the test's highest, at the 99.9th percentile (Galatzer-Levy et al., 2024).
- Against the WAIS-IV norms (mean 10, SD 3, from 2,200 adults), about two-thirds of people score between 7 and 13, so a score of 19 sits 3 SD above the human mean and 6 points above the top of that range; because 19 is the ceiling, the true margin may be larger (Block et al., 2025).
- The spread across systems on the knowledge questions was narrow: all 10 models from OpenAI, Google and Anthropic scored 18 or 19, including the small on-device Gemini Nano, whereas word-meaning subtests varied far more (Vocabulary 4 to 19, Similarities 1 to 15) (Galatzer-Levy et al., 2024).
- The comparison is treated as direct because the questions, scoring and norms were the human ones and the answer was not in view for either side, but items were typed rather than spoken, it is a single preprint, and the authors report no check on whether the test items appeared in the models' training text (what the model learned from before release) (Galatzer-Levy et al., 2024).
- Facts held in a model's training weights track how often they appear online: across 10 models and 14,000 questions, accuracy on the 4,000 least-popular facts stayed at 15 to 19% even as models grew, while retrieval, looking facts up in an external store and placing them in the conversation, let a far smaller model beat GPT-3 on those facts; no human baseline was run (Mallen et al., 2023).
- On SimpleQA, a 2024 preprint set of 4,326 obscure fact questions deliberately chosen to defeat GPT-4o, models ranged from 5.1% correct (Claude 3 Haiku) to 42.7% (o1-preview); a human checker scored 94.4% but could search the web, so this gap favours the side the test was easier for and is not counted as a comparison (Wei et al., 2024).
- When five models (GPT-3.5-Turbo, GPT-4, and three LLaMA models) did the classic category-fluency task, the semantic networks built from their answers were less interconnected and less flexible than human networks; this concerns how knowledge is organized and searched, not how much is known, and gives no score on the human scale (Wang et al., 2024).
Memory Systems · Declarative
Episodic Memory
Definition
- Episodic memory is the memory system that lets people remember specific past experiences; Tulving's mature account ties it to a sense of self, subjective time, and autonoetic consciousness (mentally re-experiencing the event), which separates it from semantic memory for general facts (Tulving, 2002).
- In animal work, where re-experiencing cannot be asked about, the behavioural criterion is memory for what happened, where, and when; scrub jays met it by searching for perishable wax worms only when little time had passed since caching, which the authors called episodic-like memory (Clayton & Dickinson, 1998).
- In personal memories, episodic details (happenings, places, perceptions, and thoughts tied to one time and place) can be scored apart from semantic details not tied to any occasion; younger adults favoured episodic details and older adults semantic ones, even after structured probing (Levine et al., 2002).
Neural Substrate
- Bilateral removal of the medial temporal lobes, including the hippocampus, caused lasting loss of memory for new experiences in patient H.M. and other surgical cases, while earlier memories and intelligence were largely spared (Scoville & Milner, 1957).
- Three patients with bilateral hippocampal damage from birth or early childhood had pronounced amnesia for everyday episodes yet reached low-average to average language, literacy, and factual knowledge, suggesting that only the episodic component depends fully on the hippocampus (Vargha-Khadem et al., 1997).
- Across imaging studies, the hippocampus and posterior parahippocampal gyrus are disproportionately active during recollection of an event with its context, and anterior parahippocampal (perirhinal) cortex during item familiarity; the hippocampus is proposed to bind items to their context (Diana et al., 2007).
- A meta-analysis of 74 fMRI studies found that activity predicting later memory centres on left inferior frontal cortex, fusiform cortex, the hippocampal formation, premotor cortex, and posterior parietal cortex (Kim, 2011).
Human Measurement
Level of Measurement: Directly Measured
- Free recall of studied lists is a standard episodic test; people tend to recall items studied close together in time one after another (the temporal contiguity effect), a benchmark seen across recognition, paired associates, and autobiographical recall and at scales from minutes to years, shown by 96–100% of individuals in large datasets such as PEERS (172 younger and 38 older adults) (Healey et al., 2019).
- A meta-analysis of 89 comparisons from 36 articles found older adults score below younger adults by Hedges' g = 0.89 (95% CI 0.75–1.03) on free recall but only 0.54 (0.37–0.72) on item recognition (Rhodes et al., 2019).
- Pooling 90 studies with 3,197 younger and 3,192 older adults, a meta-analysis found age effects larger for associative memory (source, context, temporal order, spatial location, and item pairings) than for memory of the items themselves (Old & Naveh-Benjamin, 2008).
- An earlier meta-analysis of 46 studies found that age differences in memory for the context of an event are reliably larger than for its content, making context memory the more sensitive episodic measure (Spencer & Raz, 1995).
AI Measurement
Level of Measurement: Directly Measured
AI Memory Performance: Below Human approximate provisional
- A peer-reviewed temporal-order study (ICML 2026) recruited 97 Goodreads readers who had finished The Murder of Roger Ackroyd within the previous 30 days and, without the book, asked each which of two 50-word passages came first for about 10 pairs; people were near chance for nearby passages and reliably above chance for widely separated ones (Pink et al., 2026).
- Current conversation: Llama 3.1 8B and 70B (2024) judged the same 296 passage pairs with the whole novel in their context window, meaning they could still see the book; both reached at least 70% for far-apart passages, the 8B model never differed significantly from people (all p ≥ .289), and the 70B model beat people only for passages 16,000–24,000 words apart (p = .021), so under the uneven-test rule this counts as a tie and the one above-human bin is not counted (Pink et al., 2026).
- Training, the fair condition where neither side could look back: a 2024 preprint by the same group found Llama3-8B and Mistral-7B trained on the books for five epochs no better than copies trained without them (about 51%, near chance), while 155 readers averaged 64% on 50-word passages (56% on 20-word ones), were significantly above chance beyond 4,000 words, and peaked at 76% beyond 25,000 words; this condition was dropped from the peer-reviewed version (Pink et al., 2024).
- Saved memory and spread across systems in the same preprint: four models using retrieval (RAG, meaning the system searches stored passages and pastes matches into its input) scored 55–67% at distances up to 16,500 words, overlapping the human scores with no statistical comparison, while 11 models given book excerpts in view ranged from 56% (Llama2-7B) to 95% (Mixtral-8x22B), a range that does not count against people because the text was visible (Pink et al., 2024).
- The rating is Below Human, approximate and provisional: the two results disagree, and the training test, where neither side could look back at the book, better captures episodic memory, which is remembering an experience rather than consulting a record of it, so it outranks the tie made with the book in view; that test is the weaker study (humans read one book, models nine; only two small models), and the rating rests on one research group's task and 2023–2024 open models (Pink et al., 2026).
- A 2026 preprint gave 50 Prolific adults and nine models (including GPT-5.4 and Claude Opus 4.6) a short story with questions about event order and a naturalistic story free-recall task; people averaged 7.96 of 10 and a recall similarity of 0.60, models 9.22–9.80 and 0.68–1.00, but the story was hidden from people and stayed in view for the models, so these higher scores are not counted, and a memory-limited agent built to imitate people is a simulation rather than a test of ability (Wang et al., 2026).
- A peer-reviewed benchmark with no human baseline asked models to recall events by time, place, entity, and content from a synthetic book of about 100,000 tokens; with one matching event F1 (a combined accuracy score from 0 to 1) ranged from 0.05 (o1-mini) to 0.81 (GPT-4o) with the book in view, every model fell to 0.60 or less once two or more events matched the cue, and no model recalled the complete set of events for more than 18% of multi-event questions (Huet et al., 2025).
- Saved memory in commercial assistants: across short three-to-six-session chat histories, ChatGPT's memory feature answered 57.7% (GPT-4o) and 71.1% (GPT-4o-mini) of 97 questions correctly and Coze 33.0% and 24.7%, versus 91.8% when GPT-4o read the full history directly; the benchmark reports no human baseline (Wu et al., 2025).
- A peer-reviewed NeurIPS 2024 study found that attention heads in pretrained models favour the item that followed a repeated item, a temporal contiguity and forward bias resembling human free-recall curves (PEERS, N = 171); this is a mechanistic parallel to the temporal context model, not a comparison of how much either side remembers (Ji-An et al., 2024).
- A 2025 preprint gave GPT-4 (with Mistral and Llama checks) word lists from the Deese-Roediger-McDermott paradigm and found perfect recall with no primacy or recency effect, set against published human serial-position curves; because the list stayed in view for the model, exceeding people is not counted (Cao et al., 2025).
Memory Systems · Non-Declarative
Priming
Definition
- Priming is a nonconscious form of memory in which an earlier encounter with a word or object makes later identification of it easier; perceptual priming works on the form of a stimulus, whereas conceptual priming draws on semantic memory (Tulving & Schacter, 1990).
- Structural (syntactic) priming is the tendency to repeat, or process more easily, a sentence that shares its grammatical structure with a recently experienced prime sentence, even when the words differ (Pickering & Ferreira, 2008).
- Semantic priming, where a related word speeds responses to a target, can occur between words related in meaning but not associated by habit (for example, category members), which separates it from simple word association; adding an associative link gives an extra "associative boost" (Lucas, 2000).
- What separates priming from explicit memory is the instruction, not the cue: given the same three-letter word stems, people told to complete them with the first word that comes to mind show priming, while people told to recall studied words are using conscious recollection (Graf et al., 1984).
Neural Substrate
- Patients with amnesia were impaired on free recall, cued recall and recognition but normal on word-stem completion priming, which declined at a normal rate over about two hours, so priming does not depend on the brain regions damaged in amnesia (Graf et al., 1984).
- Behavioral priming is typically accompanied by reduced activity in several cortical regions when a stimulus repeats (repetition suppression); the link between this reduction and behavioral priming is reliable in prefrontal regions but not in posterior ones (Schacter et al., 2007).
- Combining fMRI with transcranial magnetic stimulation, disrupting left frontal cortex abolished both the neural and the behavioral markers of repetition priming for object classification, while priming-related reductions in early sensory cortex were unaffected, pointing to separable conceptual and perceptual components (Wig et al., 2005).
- In fMRI, repeating a sentence structure produced adaptation in left inferior frontal gyrus (BA 45), left middle temporal gyrus (BA 21) and bilateral supplementary motor area, equally when priming ran within or between speaking and listening (Segaert et al., 2012).
Human Measurement
Level of Measurement: Directly Measured
- The standard semantic priming paradigms are speeded naming and lexical decision to a target after a related versus unrelated prime; the Semantic Priming Project collected both measures for 1,661 target words from 768 participants at four universities (Hutchison et al., 2013).
- A meta-analysis of 26 studies found reliable automatic semantic priming without word association, largely unaffected by prime duration or relatedness proportion, with smaller effects in naming than in lexical decision and smaller effects than associative priming (Lucas, 2000).
- A meta-analysis of 39 effect sizes on repetition priming found only a modest age deficit (weighted mean d = 0.30, 95% CI 0.22 to 0.39), smaller than the age deficits on recall and recognition in the same experiments (La Voie & Light, 1994).
- People who had seen pictures for 1 to 3 seconds identified fragments of those pictures better than new ones 17 years later, priming was stable over the years (r = .51), and it appeared even in people with no memory of the original session (Mitchell, 2006).
- Structural priming in production, measured by having people describe or complete a target after a prime of one structure or the other, has been pooled in a meta-analysis that identified word overlap between prime and target as its most consistent moderator (the "lexical boost") (Mahowald et al., 2016).
AI Measurement
Level of Measurement: Directly Measured
AI Memory Performance: Comparable to Humans approximate
- Run on a classic 1998 sentence-completion experiment (32 prime and target pairs, 1,000 runs each), ChatGPT and Vicuna reused the primed structure (ChatGPT 0.71 vs 0.58; Vicuna 0.81 vs 0.51), more so when prime and target shared a verb, as people do. Against the human means plotted in the paper (an effect of about 5 points with different verbs and 18 with the same verb), ChatGPT's effect (about 9 and 14) sits within the human values and Vicuna's (about 21 and 38) above them (Cai et al., 2024).
- The comparison is approximate: materials and procedure differed, the human means come from the original study with no reported spread, and no test compared people with the models. The within-range reading still holds because the models could still see the prime in the conversation, so Vicuna's larger effect cannot count in its favour, and ChatGPT, the stronger system (it matched the human pattern in 10 of 12 experiments, Vicuna in 7), sits among the human values (Cai et al., 2024).
- Seven multilingual models (XGLM, 564 million to 7.5 billion parameters, and PolyLM, 1.7 and 13 billion) were run on the materials of 12 human structural priming experiments across six languages, including priming from one language to another. All models showed priming on the single-language materials and most showed it on 8 of the 12 cross-language sets, with larger models usually showing larger effects; the authors compared only the direction of effects with people, not their size (Michaelov et al., 2023).
- Across GPT-2 Large and 7-billion-parameter Llama-2, Falcon, Mistral and Zephyr models (base and tuned versions), priming was far more lopsided than in people: sentences that strongly primed one dative structure tended to prime the other weakly or negatively (correlation -0.72 to -0.77), with only 26 to 38% priming in both directions, whereas people generally show positive priming for both. Like people, the models were primed more by rarer structures (the inverse frequency effect) (Jumelet et al., 2024).
- Using the same English dative sentences that people read during EEG recording, 13 transformer language models showed priming, but at both noun phrases after the verb, whereas people showed it only at the second. Across the full set of models, model priming scores were not good predictors of human priming effects (Sinclair et al., 2026).
- Using the materials of a human study in which solving bracketed arithmetic shifted relative-clause attachment in 36 French speakers (in eye movements, among the math-skilled), five current models (Llama 3.3 70B, Qwen3 8B, Kimi K2, gpt-oss-120b and gpt-oss-20b) showed no robust cross-domain priming, even Qwen3, which solved every equation. This one condition falls below people, though the human and model measures differ (eye tracking versus an explicit choice) (Lee et al., 2026).
- Beyond the conversation, a 2025 preprint found a training-level analogue it calls priming: after fine-tuning (further training that changes the model's stored weights) on one new fact, PaLM-2, Gemma and Llama models applied it in unrelated contexts, predictably from the fact's pre-training word probabilities, but with no human comparison. No study was found that tests priming from AI assistants' saved memory (Sun et al., 2025).
Memory Systems · Non-Declarative
Non-Associative Learning (Habituation and Sensitization)
Definition
- Habituation is a drop in a behavioral response caused by repeated stimulation that is not due to sensory adaptation, sensory fatigue, or motor fatigue; a revised consensus lists ten defining characteristics, adding long-term habituation to the original nine (Rankin et al., 2009).
- What separates habituation from simple fatigue or adaptation is that the decrement is stimulus-specific, recovers with rest (spontaneous recovery), and is reversed when a different stimulus is presented (dishabituation), criteria first set out as parametric characteristics in the 1960s (Thompson & Spencer, 1966).
- Sensitization is the strengthening of a response by strong or repeated stimulation; both it and habituation count as non-associative because they need only the stimulus itself, with no pairing of one event with another as in conditioning (Byrne & Hawkins, 2015).
- Dual-process theory holds that habituation and sensitization are separate processes set off by the same repeated stimulus, and the response people or animals show is the net result of the two (Groves & Thompson, 1970).
Neural Substrate
- In the sea slug Aplysia, habituation of the gill-withdrawal reflex was traced to weakening transmission at the synapse between sensory and motor neurons, and dishabituation to its restoration (Castellucci et al., 1970).
- Short-term habituation in Aplysia (under 30 minutes) involves presynaptic depression, whereas long-term habituation (over 24 hours) also requires postsynaptic AMPA and NMDA glutamate receptors, protein phosphatases, and calcium channels (Glanzman, 2009).
- Sensitization works through serotonin released by facilitatory interneurons, which acts via cAMP/PKA and PKC signaling at the same sensory-motor synapse; long-term sensitization additionally needs CREB-regulated gene transcription and new protein synthesis (Byrne & Hawkins, 2015).
- In humans, fMRI showed the amygdala responds more to fearful than neutral faces and that this response habituates rapidly over repeated presentations (Breiter et al., 1996).
Human Measurement
Level of Measurement: Directly Measured
- Habituation of the startle reflex and skin conductance to repeated loud tones is a standard psychophysiology paradigm; a meta-analysis pooling 25 startle studies found slower skin-conductance habituation in adults with PTSD (r = .21), one of the most robust physiological markers of the disorder (Pole, 2007).
- The infant visual habituation paradigm is widely replicated: a meta-analysis found that habituation and recognition memory scores in the first year predict IQ at ages 1 to 8, with a weighted average correlation of .36 (raw median .45) (McCall & Carriger, 1993).
- A meta-analysis of 17 study groups found the cold-shock response habituates after about four cold-water immersions, with large pooled effects such as a drop of 14 beats per minute in heart rate (d = -1.19, 145 participants) (Barwood et al., 2024).
- A meta-analysis of 47 studies that repeated the Trier Social Stress Test found robust, rapid habituation of the cortisol (HPA-axis) stress response, but less consistent habituation of autonomic and immune responses (Barthel et al., 2025).
- In 25 healthy adults scanned twice, amygdala habituation across a session was a more reliable individual measure than average amygdala activation (test-retest ICC = .53 vs .16) (Plichta et al., 2014).
AI Measurement
Level of Measurement: Indirectly Measured
AI Memory Performance: Comparison Not Possible
- No study located gives an AI system the habituation paradigm psychologists use, which checks a response decrement against the defining hallmarks (stimulus specificity, spontaneous recovery, dishabituation); nothing was found for training, the current conversation, or saved memory (stored notes an assistant can look up later), so existing work measures related effects under other names or in only one of these places (Rankin et al., 2009).
- Within the context window, meaning the text a model can see in the current conversation, GPT-2 and BART showed a sensitization-like "self-reinforcement effect": the more times a sentence had already been repeated, the more likely the model was to produce it again, a pattern studied as a text-generation flaw, not as learning (Xu et al., 2022).
- A peer-reviewed security paper tested chatbots by the name habituation: repeatedly exposing a model to a request, or escalating it in small steps, raised average jailbreak success (getting the model past its safety rules) from 0.1 to 0.5 and from 0.2 to 0.4; the abstract does not report which models were tested or how they differed, and there was no human baseline or hallmark testing (Zhang et al., 2024).
- In training, a 2026 preprint on a single small open model (1.1 billion parameters) reported "progressive adversarial sensitization": ten rounds of preference tuning raised jailbreak rates by 20%, an analogy to kindling rather than a test of the behavioral definition, with no human comparison (Cheung, 2026).
- The closest same-task comparison had 100 people and GPT-2 predict the next word of passages repeated up to four times: agreement was high on first reading (r = .87) but fell to r = .05 by the third, because GPT-2 became near-perfect after one repetition while people improved gradually; only one AI model was tested, so no spread across systems can be given (Vaidya et al., 2023).
- That comparison cannot be used for a rating: it measures memory for repeated text (closer to repetition priming) rather than a waning or growing response, and GPT-2 could still see the earlier passage while people saw words one at a time and had to remember them, so the AI's better score only favours the side the test was easier for (Vaidya et al., 2023).
- Response suppression to repeated input has been shown in AI mainly when researchers build it in: an image-recognition network given built-in suppression reproduced adaptation aftereffects seen in people, but this is sensory adaptation, which the psychology definition separates from habituation, and the component is not part of today's AI assistants (Vinken et al., 2020).
Memory Systems · Non-Declarative
Classical Conditioning (Delay and Trace)
Definition
- Classical (Pavlovian) conditioning is learning that a neutral cue, the conditioned stimulus (CS, such as a tone), predicts a significant event, the unconditioned stimulus (US, such as an air puff), so that the cue alone comes to trigger a conditioned response; in delay conditioning the CS overlaps and ends with the US, while in trace conditioning a silent gap of about 500 to 1000 ms separates them (Clark & Squire, 1998).
- What is learned is the predictive relation, not mere pairing: presenting the US when the CS is absent weakens fear conditioning, showing that CS–US contingency determines how much is learned (Rescorla, 1968).
- Delay eyeblink conditioning is treated as nondeclarative memory because it is intact in amnesia and does not depend on awareness of the CS–US relation, whereas trace conditioning depends additionally on the hippocampus and neocortex and requires awareness, placing it closer to declarative memory (Clark et al., 2002).
Neural Substrate
- The cerebellum and its brainstem sensory and motor pathways form the essential circuit for delay eyeblink conditioning, while trace eyeblink conditioning requires both the cerebellum and the hippocampus (Christian & Thompson, 2003).
- Patients with bilateral medial temporal lobe amnesia, who had shown normal delay eyeblink conditioning, produced significantly fewer conditioned responses than matched controls at trace intervals of 500, 750, and 1,000 ms (McGlinchey-Berroth et al., 1997).
- In healthy adults scanned with fMRI while learning delay and trace eyeblink conditioning in parallel, cerebellar activation was comparable for the two forms, but hippocampal activation was greater for trace conditioning (Cheng et al., 2008).
- A meta-analysis of 27 human fear-conditioning fMRI studies (677 participants) found a consistent network of activation resembling the central autonomic-interoceptive network, plus coordinated deactivations that may reflect processing of safety signals (Fullana et al., 2016).
Human Measurement
Level of Measurement: Directly Measured
- The standard paradigms are eyeblink conditioning (tone CS, air-puff US, scored as the percentage of anticipatory blinks) and differential fear conditioning (a CS+ paired with shock versus an unpaired CS−, scored by skin conductance and ratings); amnesic patients and healthy volunteers tested on two delay and two trace versions showed that awareness was unrelated to delay conditioning but a prerequisite for trace conditioning (Clark & Squire, 1998).
- Across four replications of differential delay eyeblink conditioning (12 to 25 adults each, 120 trials), group means ranged from about 3 to 25 percentage points more conditioned responses to the CS+ than the CS−, with aware and unaware participants conditioning equally (Smith et al., 2005).
- A 2025 meta-analysis pooling 77 fear-conditioning studies (2,052 patients with anxiety- or stress-related disorders and 3,258 healthy controls) found reliable group differences in responses to safety cues and during extinction, but no difference in CS+/CS− discrimination, updating an earlier meta-analysis of 44 studies (Kausche et al., 2025).
- When 107 adults completed the same fear-conditioning paradigm and 71 repeated it six months later, conditioning was robust at the group level, but individual-level test-retest reliability of skin conductance responses was limited (intraclass correlations about 0.03 to 0.61 depending on analysis choices) (Klingelhöfer-Jens et al., 2022).
- A systematic review and meta-analysis of 41 studies found no convincing evidence that fear conditioning occurs without awareness of the CS–US contingency, with higher-quality studies less likely to report unaware conditioning and signs of publication bias (Mertens & Engelhard, 2020).
AI Measurement
Level of Measurement: Indirectly Measured
AI Memory Performance: Below Human provisional
- Only one study found tests AI on classical conditioning by name, a peer-reviewed 2026 benchmark covering the current conversation and add-on saved memory; it uses a text-based analogue rather than the eyeblink or fear paradigm, and no study was found that tests conditioning acquired during training, so the ability is only partly measured (Qin et al., 2026).
- Each of the 100 conditioning items pairs a cue with a bad outcome over four conversational turns (for example, requests for an "http" link each followed by an "insecure protocol" warning), adds two unrelated turns, then brings the cue back; an item counts as correct only if the first reply avoids or warns about the cue, judged against a rubric by a separate AI model (Qin et al., 2026).
- Across 17 large language models (averaged over three runs), conditioning scores ranged from 35.7% (Qwen-2.5-7B) to 69.7% for the best system (DeepSeek-R1), with GPT-5 at 64.0% and Claude-4.1-opus at 41.7%; avoidance-type items fared worst, averaging 17.6% versus 75.0% for preference-type items (Qin et al., 2026).
- The human baseline was five computer science Ph.D. students who completed the same items under the same protocol, scored with the same rubric by two other Ph.D. students (100% agreement); all five scored 100%, so the best AI fell about 30 points below the entire human range, though the human sample is tiny, expert, at ceiling, and scored by people rather than the AI judge (Qin et al., 2026).
- The test was, if anything, easier for the AI: the models could still see the whole short dialogue (about 500 tokens) in their context window, meaning the text of the current conversation, and the paper does not say whether people could reread it, so the below-human result goes against the AI's advantage and counts (Qin et al., 2026).
- Adding saved memory did not close the gap: memory-augmented agents scored 63% (MEM1), 25% (MemGPT), and 22% (MemAgent) on conditioning, and even when researchers manually stored the exact cue–outcome pairing in an external memory tool (Mem0), the best score was 76% (Qin et al., 2026).
- A 2025 preprint shows that the update rules of linear-attention transformer layers are mathematically identical to classic conditioning models (Hebbian contiguity and Rescorla–Wagner error correction) and reproduce Kamin blocking in small trained networks, but it did not test deployed language models or compare them with people (Qiao, 2025).
Memory Systems · Non-Declarative
Instrumental Learning (Action–Outcome and Habit)
Definition
- Instrumental learning is learning from the consequences of one's own actions, and it comes in two forms: goal-directed actions, which depend on knowing what an action produces and how much that outcome is currently wanted, and habits, which are stimulus–response links that run regardless of the outcome's current value (Dickinson, 1985).
- Goal-directed action rests on two kinds of learning, contingency learning (which action leads to which outcome) and incentive learning (how valuable that outcome is), whereas a separate stimulus–response habit mechanism does not track outcome value (Balleine & Dickinson, 1998).
- What marks instrumental learning is that the learner's own action, not merely a cue that precedes an outcome, becomes linked to the consequence; decision tasks in humans and rodents show separate contributions from action–outcome learning and stimulus–response learning (Balleine & O'Doherty, 2010).
- Computational accounts recast the two forms as model-based learning, which plans using a map of which actions lead where, and model-free learning, which simply repeats what was rewarded, though whether these map one-to-one onto goals and habits is still debated (Daw et al., 2011).
Neural Substrate
- Rodent and human studies trace action–outcome learning and stimulus–response habit learning to distinct corticostriatal networks that appear homologous across species, with early lesion work pointing to the rat prelimbic prefrontal cortex for learning which action produces which outcome (Balleine & O'Doherty, 2010).
- In an fMRI outcome-devaluation study, orbitofrontal cortex activity was strongly modulated when people chose an action whose outcome had been devalued compared with one that had not (Valentin et al., 2007).
- Extensive training in humans reduced sensitivity to outcome devaluation and increased cue-related activity in the right posterior putamen, the human counterpart of the rodent dorsolateral striatum implicated in habits (Tricomi et al., 2009).
- During human instrumental learning, drugs that boost or reduce dopamine function changed the size of striatal reward prediction errors and, with them, how often people chose the rewarding option (Pessiglione et al., 2006).
Human Measurement
Level of Measurement: Directly Measured
- The standard test is outcome devaluation: people learn actions that earn particular outcomes, one outcome is devalued (for example by eating it to satiety), and continued responding for the devalued outcome is taken as habitual; with extensive training, people in one fMRI study became less sensitive to devaluation (Tricomi et al., 2009).
- Habit induction in healthy adults has proved hard to replicate: across five experiments (N = 43 to 74 each; avoidance, slips-of-action and free-operant versions) much longer training did not make responding less sensitive to devaluation, with Bayes factors up to about 4 favouring no effect (de Wit et al., 2018).
- A preregistered replication with a stronger devaluation procedure found most participants were goal-directed after both short and extended training, and that apparent habits came mainly from participants for whom the devaluation had not worked (Vázquez-Millán et al., 2026).
- The two-step task measures the goal-directed (model-based) side; in online samples of 548 and 1,413 adults, a one-standard-deviation increase in compulsive symptoms predicted a 14% to 17% reduction in the model-based index, and the second sample replicated the first (Gillan et al., 2016).
- When the two-step task is explained in detail, healthy adults behave mostly model-based with little model-free influence, showing that task instructions shift the measured balance (Feher da Silva & Hare, 2020).
AI Measurement
Level of Measurement: Indirectly Measured
AI Memory Performance: Comparable to Humans approximate provisional
- In training, language models are shaped by reinforcement learning from human feedback, meaning their internal settings are adjusted according to how people rated the model's own outputs; this is learning from consequences, but it is judged by output quality, not by instrumental-learning experiments (Ouyang et al., 2022).
- CogBench ran 35 language models in conversation on the probabilistic instrumental learning task from Lefebvre et al. (2017), four interleaved slot-machine pairs over 96 trials, scoring money earned on a scale where 0 is a random chooser and 1 is the average human participant; all seven established models in its main figure reached human level on this task, and GPT-4 and Claude-1 reached it on five of six tasks, while every one of the seven reached it on at least three (Coda-Forno et al., 2024).
- The human results behind that scale are documented: 50 healthy adults in an fMRI experiment and 35 more in a behavioural one chose repeatedly between pairs of symbols that paid out with 25% or 75% probability, learning from wins and losses (Lefebvre et al., 2017).
- The comparison is approximate and uneven: the task and money scale were shared, but every past choice and reward stayed written in the model's prompt while people had to learn and remember them, and CogBench reports only the human average, with no spread and no statistical test against people; GPT-4 therefore matched the human average on a test that favoured it, a tie that counts as Comparable and is provisional (Coda-Forno et al., 2024).
- Systems differ widely beyond CogBench's leaders: of GPT-3.5, GPT-4 and Llama 2 tested on bandit problems, only GPT-4 with step-by-step reasoning and an externally summarised interaction history explored satisfactorily (Krishnamurthy et al., 2024).
- GPT-3 on the horizon bandit task was indistinguishable from people with short horizons (p = .97) but earned significantly lower regret overall (M = 2.72 vs 3.24, human SD = 10.26), with past outcomes shown in the prompt; this result is not used because the human data come from a conference abstract (Zaller et al., 2021) whose sample size and procedure the paper does not report (Binz & Schulz, 2023).
- In saved memory, agents such as Reflexion store written lessons from task feedback to guide later attempts, but no study has tested this, or any AI system, on outcome devaluation or habit formation, so the rating rests on reward learning within a single conversation (Shinn et al., 2023).
Memory Systems · Non-Declarative
Procedural Memory
Definition
- Memory divides into a system accessible to conscious recollection (declarative memory) and non-declarative systems that are expressed in performance; procedural memory for skills and habits is a non-declarative system supported by the neostriatum and cerebellum rather than the hippocampus (Squire, 2004).
- Amnesic patients learned a mirror-reading skill as fast as matched controls and kept it for at least 3 months, separating rule- or procedure-based "knowing how" from data-based "knowing that" (Cohen & Squire, 1980).
- Implicit learning of rule systems was established with artificial grammars: people exposed to letter strings generated by a hidden finite-state grammar later classify new strings by that grammar without being able to state its rules (Reber, 1967).
- Implicit learning scores on four tasks (artificial grammar, probabilistic classification, serial response, implicit category learning) were unrelated to IQ, working memory or explicit learning ability, distinguishing this ability from explicit, effortful learning (Kalra et al., 2019).
Neural Substrate
- In a double dissociation, patients with Parkinson's disease failed to learn a probabilistic classification (weather prediction) task yet remembered the training episode, while amnesic patients learned the task normally but forgot the episode, implicating the neostriatum (caudate and putamen) in gradual habit learning (Knowlton et al., 1996).
- Lesion and pharmacological work in rats and monkeys, and patient and imaging work in humans, tie the dorsal striatum to incrementally acquired stimulus-response habits, a system that can compete with the hippocampal declarative system during learning (Packard & Knowlton, 2002).
- A neuroimaging meta-analysis of serial reaction time studies found that, once visual and motor factors were controlled, sequence learning consistently activated only the basal ganglia (caudate, putamen and globus pallidus); premotor and cerebellar activity reflected other aspects of the task (Janacsek et al., 2020).
- Across 70 motor learning experiments, activation converged on a cortical-subcortical network including dorsal premotor, supplementary motor and primary motor cortex, thalamus, putamen and cerebellum, with left dorsal premotor cortex active in every analysis (Hardwick et al., 2013).
Human Measurement
Level of Measurement: Directly Measured
- The serial reaction time task, in which people respond to lights or targets that secretly follow a repeating sequence and become faster on sequenced than random trials, was introduced by Nissen and Bullemer and is now the most widely used paradigm for procedural sequence learning (Nissen & Bullemer, 1987).
- A meta-analysis of 7 studies (719 participants, mean age 20.8) confirmed that the serial reaction time learning effect is robust at the group level but has poor test-retest reliability for individuals (r < .40), with better within-session split-half reliability (.66) (Oliveira et al., 2023).
- Meta-analyses of motor sequence learning (34 articles, 88 groups, 1,296 participants) replicated a large performance gain after sleep and a smaller gain after wakefulness, but once methodological confounds were accounted for there was no evidence that sleep enhances the skill (Pan & Rickard, 2015).
- Pooling 8 serial reaction time studies (186 children with specific language impairment, 203 typically developing peers) gave a small but significant procedural learning deficit, mean effect size .328 (95% CI .071 to .584) (Lum et al., 2014).
- Healthy young adults retested about a week apart showed medium test-retest reliability on probabilistic classification, serial response and implicit category learning tasks, indicating stable individual differences in implicit learning (Kalra et al., 2019).
AI Measurement
Level of Measurement: Directly Measured
AI Memory Performance: Below Human provisional
- ImplicitMemBench tests large language models on procedural memory by name: 100 items give a new rule with 1 to 3 examples (e.g., a custom arithmetic operator or reversed tool-call order), then 15 turns of misleading conversation, then score only the model's first attempt on new probes; unlike human procedural tasks, the rule is stated outright and learned in one exposure rather than through practice (Qin et al., 2026).
- Five computer science PhD students completed the same items under the same protocol and scoring rubric and all scored 100%; the best model, Claude-4.1-opus, averaged 76.67% over three runs, about 23 points below the human range, even though the rule was still in the model's context window (what it can see in the current conversation) (Qin et al., 2026).
- Across 17 models, procedural memory scores ranged from 46.67% (LLaMA-3.1-8B) to 76.67%, with eight models above 70%; the human baseline is small, unrepresentative and at ceiling, so the size of the gap is uncertain but its direction is not (Qin et al., 2026).
- Saved memory did not close the gap: storing the key rule by hand in an external memory store (Mem0) moved DeepSeek-R1 only from 76.33% to 77.00%, and among memory-augmented agents MemAgent rose from 51% to 60% while MEM1 fell from 50.67% to 27% and MemGPT from 50% to 44% (Qin et al., 2026).
- A 2025 preprint ran gpt-4o and o3-mini on three classic artificial-language experiments with published human data: on a finite-state grammar task (15 people per grammar) both models improved across blocks and found the simpler grammar easier, as people did, but on a morphosyntax task neither model improved over trials, and on a morphology task only o3-mini matched human regularization (Ma & Xu, 2025).
- That comparison is approximate and does not set the rating: models got shorter versions of the tasks (e.g., 120 rather than 480 sentences), could still see the training items in the conversation while people had to remember them, and results were compared as learning patterns rather than on a common accuracy scale (Ma & Xu, 2025).
- Outside comparisons with people, AI agents (systems that act step by step toward a goal) can now save procedural memory, distilling past attempts into step-by-step instructions and scripts; this raised task success on TravelPlanner and ALFWorld, and the saved memory also improved weaker models, but no human baseline was run (Fang et al., 2026).
Representational Structures · Schemas
Schema Formation
Definition
- Schemas are superordinate knowledge structures that capture what many separate experiences have in common, and once formed they shape how new events are perceived, interpreted and remembered (Gilboa & Marlatte, 2017).
- A memory schema is defined by four necessary features: an associative network structure, a basis in multiple episodes, a lack of unit detail, and adaptability; the multiple-episode basis and loss of detail separate it from an episodic memory of a single event (Ghosh & Gilboa, 2014).
Neural Substrate
- Rats that had built a schema of flavour-place associations over weeks of training learned new associations that fit it in a single trial, and those new memories rapidly became independent of the hippocampus, showing that an existing schema speeds consolidation into the neocortex (Tse et al., 2007).
- In a lesion study, all four patients with ventromedial prefrontal cortex damage and a history of confabulation were impaired at judging whether words belonged to a schema and at rejecting related lures, while other patients performed like healthy controls (Ghosh et al., 2014).
- Model-based fMRI during concept learning found brain signals tracking an abstract category prototype in the ventromedial prefrontal cortex and anterior hippocampus, with no significant signals tracking individual stored examples in those regions (Bowman & Zeithamova, 2018).
- Hippocampal and medial prefrontal activity patterns came to represent the features that related memories shared, but only a week after learning, suggesting that consolidation builds overlap across related experiences (Tompary & Davachi, 2017).
Human Measurement
Level of Measurement: Directly Measured
- The prototype-distortion paradigm tests schema formation directly: people study distorted versions of a pattern they never see, then classify new items, including the unseen prototype, to show what they abstracted (Posner & Keele, 1968).
- In a modern version with 29 young adults, accuracy was 86.2% (SD 10.9) on trained items and 80.5% (SD 7.9) on new items, new-item accuracy rose the closer an item was to the never-seen prototype (F(1,28) = 67.05), and almost three-quarters of participants were best fit by a prototype model (Bowman & Zeithamova, 2018).
- The schema-induction paradigm has people study two or more problems or stories that share an underlying principle and then measures transfer of that principle to a new, surface-different problem (Gick & Holyoak, 1983).
- A meta-analysis of 57 experiments (336 tests) found that comparing multiple cases, which promotes abstraction of their shared structure, improved learning over studying single or sequential cases, d = 0.50, 95% CI [0.44, 0.56] (Alfieri et al., 2013).
- Gains were larger when learners were asked to find similarities between cases, when a principle was given after the comparison, with perceptual content, and when testing was immediate (Alfieri et al., 2013).
AI Measurement
Level of Measurement: Indirectly Measured
AI Memory Performance: Below Human approximate provisional
- Training: five pretrained language models (from BERT to MPT) almost always rated possible events as more likely than impossible ones across 1,215 sentence pairs but were less consistent for likely versus unlikely events, showing that generalized event knowledge partly emerges from training text; no human memory comparison was made (Kauf et al., 2023).
- Saved memory: Agent Workflow Memory has a GPT-4-based web agent distil reusable routines from its past tasks and store them for later use, raising task success on the WebArena benchmark from 23.5% to 35.5%; this is schema-like induction under another name, with no human comparison (Wang et al., 2024).
- Saved memory: generative agents periodically condense stored observations into higher-level 'reflections' that guide later behaviour, and removing reflection made behaviour less believable in ablation tests, but the abstraction itself was not scored against people (Park et al., 2023).
- Conversation: a 2026 preprint ran four current AI models on OpenLock, a causal-structure transfer task first used with 80 people, giving models a written description of a previously solved environment in their context window (what the model can see in the current conversation); no study was found giving AI a classic schema-formation paradigm such as prototype distortion or the Tse schema task (Xiang et al., 2026).
- People who had experienced the same causal structure before cut attempts from 20.66 (SD 9.09) to 13.85 (SD 10.00), d = 0.71, and cut attempts to their first solution from 8.97 to 4.99, whereas the best model, Gemini-3-Flash, improved only from 9.08 to 7.33 attempts and no model significantly cut first-solution attempts (Xiang et al., 2026).
- Transfer benefit varied widely across systems: 33% for people versus 19.3% for Gemini-3-Flash, 17.9% for Claude-4.5-Sonnet, 10.5% for DeepSeek-V3.2 and -10.5% (worse after prior experience) for GPT-5.2 (Xiang et al., 2026).
- The comparison is approximate: people played a visual version in 2018 and had to remember earlier environments, while models played a text version with the prior solutions still visible and had lower baseline attempt counts; because the uneven test favoured the models yet they still showed less use of prior structure, the below-human result counts, though it rests on one preprint with no direct statistical test between people and models (Xiang et al., 2026).
Representational Structures · Schemas
Schema Application
Definition
- Schema application is remembering as reconstruction: when people recall unfamiliar material, such as a folk tale from another culture, they rationalize and reshape it to fit the knowledge frameworks they already hold (Bartlett, 1932).
- A memory schema is an associative knowledge structure abstracted from many episodes, lacking the detail of any one of them and adaptable to new input; this separates it from memory for a single episode and from isolated semantic facts (Ghosh & Gilboa, 2014).
- The key distinction is between the schema itself and its instantiation: applying an activated schema template to a specific new situation shapes perception, encoding and retrieval, and can either enhance or distort what is remembered (Gilboa & Marlatte, 2017).
Neural Substrate
- In rats that had learned a spatial schema of flavour-place pairings, new pairings were learned in a single trial and quickly became independent of the hippocampus, showing that an existing schema speeds the integration of new information (Tse et al., 2007).
- Reviews of imaging and lesion work propose that medial prefrontal cortex supports encoding and retrieval of schema-congruent information, while the medial temporal lobe handles novel or incongruent information; a later review adds angular gyrus and associative cortices to this vmPFC-hippocampus network (van Kesteren et al., 2012).
- Patients with ventromedial prefrontal cortex (vmPFC) damage (n = 7) falsely recalled fewer schema-related critical words in the DRM word-list task than healthy comparisons (n = 14), 36.5% vs 52.4%, while their true recall was normal (Warren et al., 2014).
- Healthy adults (n = 12) showed more true and false recognition for objects studied in schema-congruent contexts, but patients with vmPFC damage (n = 6) showed no congruency effect despite normal overall recognition (Spalding et al., 2015).
Human Measurement
Level of Measurement: Directly Measured
- A close replication of Bartlett's repeated-reproduction procedure with "The War of the Ghosts" (tests at 15 minutes, 1 week and 6 months) found that people introduced rationalizations and distortions, that the proportion of distorted material grew with delay, and that new propositions were imported at long delays (Bergman & Roediger, 1999).
- In the room-schema paradigm introduced by Brewer and Treyens (1981), people who viewed a graduate student's office later gave false "remember" judgements to schema-expected objects that were not there, and these increased after a 48-hour delay, while atypical objects that were present were especially well recollected (Lampinen et al., 2001).
- A meta-analysis of 60 studies found a slight overall memory advantage for schema-inconsistent information, but the direction depends on the measure: uncorrected recognition favours schema-consistent items (schema-guided guessing), while corrected recognition and recall favour inconsistent items (Rojahn & Pettigrew, 1992).
- Normative data for 36 DRM lists, in which people study associates of an unpresented word, show a wide human range: 60% or more of participants falsely recalled the critical word for the most effective lists (with false recognition above 80%), versus 10% false recall and 27% false recognition for the weakest list (Stadler et al., 1999).
AI Measurement
Level of Measurement: Directly Measured
AI Memory Performance: Not Yet Compared
- Using the story from a human serial-reproduction study of gender stereotypes (Kashima, 2000), five preregistered experiments had ChatGPT (January 2023 version) repeatedly retell the story in chains; like people, it passed on stereotype-consistent content more than inconsistent content (beta = 0.058, p < .01), but unlike people it did so from the first retelling, where human participants had favoured inconsistent details (Acerbi & Stubbersfield, 2023).
- That comparison is uneven and off-scale: the model was told to summarize a text it could still see, whereas people retold the story from memory, and results were reported as regression slopes for one AI system with no matching human spread, so no rating can be drawn from it (Acerbi & Stubbersfield, 2023).
- A 2025 preprint ran the DRM false-memory task on three language models with the word lists kept in the context window, meaning the text the model can see in the current conversation; GPT-4 recognized all studied words, falsely recognized 5.3% of unrelated words and 11.4% of critical lures, while Mistral-7B and Llama-3-8B falsely recognized essentially no lures (Cao et al., 2025).
- GPT-4's 11.4% lure rate falls below the human range of roughly 27% to over 80% false recognition, but because the model could still see the lists, fewer false memories is exactly what its advantage predicts, so under the uneven-test rule this result does not count as a comparison (Stadler et al., 1999).
- For information held in the model's trained weights, researchers fine-tuned Mistral-7B on Bartlett's "War of the Ghosts" after giving it background training on one of five Wikipedia topics; its recalls drifted toward that background topic with topic-related intrusions, matching Bartlett's and Bergman and Roediger's findings only qualitatively, in a research model rather than a deployed assistant (Spens & Burgess, 2026).
- For retrieved (saved) memory, a 2024 preprint tested six AI systems on 1,200+ questions paired with documents containing altered answers; systems wrongly adopted the altered document 15.7% (Claude Opus) to 31.3% (GPT-3.5) of the time and did so less often the further the alteration departed from their prior knowledge, a knowledge-conflict measure with no human baseline (Wu et al., 2024).
- Not yet compared: the classic schema paradigms have been given to AI, but the only same-scale result (DRM) favours the side that could see the material, the story-retelling and Bartlett studies report only qualitative or regression-slope matches, and the retrieval study has no human data (Cao et al., 2025).
Representational Structures · Person Models
Autobiographical Memory
Definition
- Autobiographical memories are temporary mental constructions assembled within a self-memory system, in which a knowledge base about one's own life is accessed under the control of the current goals of the "working self", and that knowledge base in turn grounds those goals (Conway & Pleydell-Pearce, 2000).
- Retrieval is a hierarchical search from general knowledge toward a single event; many emotionally disturbed people stop early and summarize categories of events (overgeneral memory) instead of recalling a specific episode (Williams et al., 2007).
- It spans both episodic recollection of specific personal events and personal semantics, meaning facts about one's own past such as where one lived, which is idiosyncratic like episodic memory but abstract like semantic memory (Renoult et al., 2012).
Neural Substrate
- A meta-analysis of 24 imaging studies identified a core, left-lateralized network for autobiographical memory: medial and ventrolateral prefrontal cortex, medial and lateral temporal cortex, retrosplenial/posterior cingulate cortex, temporoparietal junction and cerebellum (Svoboda et al., 2006).
- Quantitative (ALE) meta-analyses showed that autobiographical memory shares a core network with prospection, navigation, theory of mind and the default mode, with extensive overlap between autobiographical memory and theory of mind (Spreng et al., 2009).
- Patient K.C., with large bilateral hippocampal lesions, kept much general knowledge including knowledge about himself but could not recollect any personally experienced event from his whole life, dissociating episodic from personal semantic memory (Rosenbaum et al., 2005).
- An ALE meta-analysis found the left hippocampus and posterior cingulate active both when a memory is first constructed and while it is elaborated, with ventromedial prefrontal cortex, left angular gyrus, right hippocampus and precuneus adding to construction (Daviddi et al., 2023).
Human Measurement
Level of Measurement: Directly Measured
- The Autobiographical Interview, built for this ability, has people recall events from five life periods and scores internal (episodic) versus external (largely semantic) details; younger adults favored episodic details while older adults favored semantic details, even after structured probing (Levine et al., 2002).
- The Autobiographical Interview has been used in more than 200 studies, and automated scoring correlated strongly with manual internal- and external-detail counts in four of five datasets (van Genugten & Schacter, 2024).
- In the cue-word method, 70-year-olds showed childhood amnesia, a power-function retention curve for the most recent 10 years, and a reminiscence bump of extra memories from ages 10 to 30, holding for individuals and for both 124 and 921 cued memories (Rubin & Schulkind, 1997).
- A systematic review of 68 studies found the reminiscence bump consistently, at about ages 10 to 30 for important memories and 5 to 30 for word-cued memories (Munawar et al., 2018).
- Meta-analysis of the Autobiographical Memory Test showed people with remitted depression recalled fewer specific memories than never-depressed people (k = 15, g = -0.31, 95% CI -0.54 to -0.09) and more categoric memories (k = 9, g = 0.25) (Hallford et al., 2022).
AI Measurement
Level of Measurement: Indirectly Measured
AI Memory Performance: Comparison Not Possible
- Training holds no record of the model's own experiences: ChatGPT (versions 3.5 and 4.0) asked for self-defining memories produced invented, redemptive stories, and readers identified human narratives correctly about 71% of the time and ChatGPT narratives about 60% of the time (Mansfield et al., 2026).
- When GPT-4, Claude Haiku and Claude Opus stood in as simulated participants on the Emotional Recall Task alongside 1,200 people, GPT-4 came closest to human patterns but with less variability; the authors note that LLMs lack autobiographical memory, so only word associations were compared (Carini et al., 2025).
- In saved memory, Generative Agents keep a time-stamped "memory stream" of their own experiences plus reflections and plans, a structure resembling a self organized around goals; in interviews on self-knowledge and memory, the full system was rated more believable (TrueSkill 29.89) than crowdworker-written answers (22.95) or a no-memory version (21.21), though it embellished recalled events (Park et al., 2023).
- That human comparison does not measure autobiographical memory: the 25 crowdworkers watched a replay of the agent's life and inspected its memory store while role-playing it, so they reported someone else's recorded past, and 100 raters scored believability rather than accuracy (Park et al., 2023).
- On LoCoMo, question answering over conversations averaging about 600 turns between agents given personas and timelines of life events, the best model, GPT-4-turbo, scored 51.6 F1 (a 0-100 answer-overlap score) against 18.4 to 42.8 for other models and a reported human score of 87.9; the paper does not report how many people were tested or how, and the task is reading another's conversation rather than remembering one's own life (Maharana et al., 2024).
- Commercial assistants with saved memory recalling a user's past chats (3 to 6 sessions) answered 24.7% to 71.1% of 97 questions correctly (best: ChatGPT with GPT-4o-mini), versus 91.8% when GPT-4o could read the full history in its context window, meaning the text it can see in the current conversation; no human baseline was run (Wu et al., 2025).
- Comparison is not possible: a 2026 preprint built an explicitly Conway-style store of life periods, general events and specific events for AI personas but judged role-play realism rather than memory, and no study located gives AI a human autobiographical paradigm (Autobiographical Interview, cue words, Autobiographical Memory Test) about its own past with documented human results on a common scale (Tang & Guo, 2026).
Representational Structures · Person Models
Partner Model
Definition
- When two people refer repeatedly to the same thing, they settle on a shared way of describing it, a conceptual pact, that belongs to that particular partnership and is reused in later references even when a simpler term would do (Brennan & Clark, 1996).
- A partner model holds partner-specific common ground, built from interaction with one person, alongside community-wide conventions that serve as default expectations for strangers; what is learned with one partner is only partly carried over to the next (Hawkins et al., 2023).
- What is shared with a particular person appears to be retrieved through ordinary memory processes, with the experience of learning something together acting as a strong cue that the partner knows it, and such partner-specific effects held even when a simpler category cue was available (Gorman et al., 2013).
- In social psychology, person memory stores two kinds of information about an individual, the specific behaviors they have shown and more abstract trait dispositions, and both shape later recall and judgments about that person (Srull & Wyer, 1989).
Neural Substrate
- Patients with hippocampal amnesia, who cannot learn arbitrary paired associates, still developed shared, simplified labels for abstract tangram shapes with a partner at the same rate as healthy comparison participants, and the labels endured long after the task (Duff et al., 2006).
- Eye-tracking showed that amnesic patients used common ground successfully when the shared information was immediately available but failed dramatically when it had to be retrieved from memory, pointing to a role for hippocampal declarative memory in maintaining common ground (Rubin et al., 2011).
- In three-party conversations, people with bilateral hippocampal damage still tailored descriptions to which partner already knew which images, but less than matched comparison participants, especially when they had to switch rapidly between what each partner knew (Yoon et al., 2024).
- In an fMRI study, which of four learned individuals a participant was imagining could be decoded from medial prefrontal cortex activity, while lateral temporal and posterior cingulate cortex tracked the individuals' traits, consistent with a stored personality model used to predict behavior (Hassabis et al., 2014).
Human Measurement
Level of Measurement: Directly Measured
- The standard test is a repeated referential communication game followed by a partner switch: after a participant and one partner build shared names for pictures, either the same partner or a new one uses the old name or a new one; in interactive eye-tracking experiments, listeners interpreted an established name faster when it came from the partner who had helped establish it (Brown-Schmidt, 2009).
- A meta-analysis re-analysing 10 eye-tracking experiments found an early, brief same-partner advantage for kept names (from about 200 ms, peaking at 450 ms, gone by 700 ms), a later different-partner advantage when a name was abandoned (reliable from 800 ms), and a partner-specificity index averaging 55% (range 44 to 67%), meaning roughly half of the benefit of a shared name is tied to the particular partner (Kronmüller & Barr, 2015).
- In an online study of 92 adults each playing tangram games with three partners in turn, listener accuracy dropped at each partner switch (b = -1.56, p < .05) and speakers' descriptions grew longer again (b = 0.43, p < .001), yet first exchanges with each new partner improved (b = 0.57, p < .01), showing partner-specific memory with partial transfer (Hawkins et al., 2020).
- Healthy adults, and patients with amnesia, gave brief descriptions of tangrams to a familiar partner and longer, fuller ones to a new partner, replicating the partner-switch effect in speaking as well as in listening (Yoon et al., 2017).
- The partner model lasts beyond one session: in two online experiments, the strength of the common ground a pair built predicted verbatim recall of their picture descriptions about a week later, and pairs of friends, who share personal history, described pictures more efficiently than pairs of strangers (Nault et al., 2023).
AI Measurement
Level of Measurement: Indirectly Measured
AI Memory Performance: Below Human approximate
- No published study gives an AI system the human partner-switch test across separate encounters, or compares with people an AI that keeps separate, lasting models of several individuals; AI memory benchmarks remain almost entirely one-user-at-a-time, so the evidence below covers pieces of the ability in different places: adapting to one partner within a conversation the model can still see, saved memory of one user, and memory for several people without any human baseline (Yang et al., 2026).
- Within one conversation, five multimodal AI models (models that read images as well as text) replayed 54 human-human tangram-style games from 108 online participants, with the whole game history kept in their context window, meaning what they can see in the current conversation; without explicit instruction GPT-4, Gemini 1.0 Pro Vision and Claude 3 Opus (2023-2024) made their descriptions longer rather than shorter across rounds, IDEFICS and LLaVA repeated themselves, and as listeners GPT-4 reached 88.9 to 92.5% against about 99.5% for humans while IDEFICS fell to chance (Hua & Artzi, 2024).
- In a live experiment with 103 adults (32 human pairs, 39 human-AI pairs, plus 18 AI-AI pairs) matching objects over four rounds, human pairs rose from 80% to over 90% accuracy while using fewer words each round (slope -74.9 words); GPT-5.2 (2025), given the full dialogue history, fell in accuracy from 90% when paired with itself (slope -5.3 points per round) and from 85% when directing a human (slope -15.3), and its word counts did not fall (Zeng et al., 2026).
- The only study that isolates partner-specificity (a 2026 preprint) compared real pairs with pseudo-pairs stitched together from different pairs' games: in 42 human pairs from a published tangram corpus, shared labels were far more common in real than pseudo pairs (odds ratio 6.85) and descriptions shrank to about 30% of their first length, whereas for GPT-5 (2025) agents, which could see the full history, label sharing was near ceiling with or without a real partner (odds ratio 1.26, 95% CI 0.99 to 1.61) because their long descriptions reused the same stock words, and descriptions stayed near 90% of their first length (Wang et al., 2026).
- Why the comparison is approximate but still clear: these games test building a partner-specific way of talking within one encounter rather than a lasting model reused across encounters, and the AI could still read the whole history while people relied on memory; because that advantage favoured the AI and it still fell short of people on the signatures of a partner model, the result counts as Below Human (Wang et al., 2026).
- Spread across systems: in a 2026 preprint GPT-5 pairs reached 99% accuracy against 93% for human pairs, but their descriptions barely shortened (58 to 56 words, against 20 to 6 for humans) and GPT-5 paired with people stayed below human pairs in accuracy and word overlap even when prompted to sound humanlike; explicit instructions to shorten (Hua & Artzi, 2024) or extra training (Vaduguru et al., 2025, up to 41% shorter messages) narrow the length gap, but prompted models kept changing their words, and the higher AI-AI accuracy seen with the history visible is not counted (Jones et al., 2026).
- Across sessions, saved memory of a single user has been tested only without human baselines: answering questions about a user's earlier chats, ChatGPT's memory feature scored 57.7% and Coze 33.0% (both on GPT-4o, 2024) against 91.8% when the same model read the chat history directly; related benchmarks of a user's changing preferences (PersonaMem, PrefEval) likewise report no human scores, so none of this can be placed on a human scale (Wu et al., 2025).
- Keeping separate records for several people is also unmatched by human data: in a 2026 preprint on group chats, eight memory methods averaged 20.6% to 46.0%, a simple keyword search scored 43.2%, and none exceeded 27.1% on keeping each person's changing information up to date; FANToM (Kim et al., 2023), where GPT-4 trailed humans on who knows what within one conversation, tests theory of mind rather than a lasting person model and is not used for this rating (Yang et al., 2026).
Memory Processes · Acquisition and Transformation
Consolidation
Definition
- Consolidation is the hypothesis, first proposed by Müller and Pilzecker a century ago, that new memories are fragile at first and become stable only gradually through time-dependent processes after learning (McGaugh, 2000).
- The term covers two processes: synaptic consolidation, completed within minutes to hours in every memory system studied, and systems consolidation, a slower reorganization in which memories that first depend on the hippocampus may become hippocampus-independent (Dudai, 2004).
- Consolidation is distinguished from reconsolidation: a memory that is already stable can become labile again when it is retrieved and must be re-stabilized, which classic consolidation theory did not predict (Nader et al., 2000).
- The active systems consolidation account holds that sleep, especially slow-wave sleep, reactivates new memories so they are redistributed to long-term cortical stores, selectively and with qualitative change, which separates it from passive protection against interference (Born & Wilhelm, 2012).
Neural Substrate
- Hippocampal damage produces temporally graded retrograde amnesia, impairing recently formed memories while typically sparing remote ones, the early evidence that the hippocampus guides a gradual reorganization of memory in neocortex (Squire et al., 2015).
- In rats, hippocampal place cells that fired together during a spatial task showed increased co-firing during the slow-wave sleep that followed, a replay of waking experience during sleep (Wilson & McNaughton, 1994).
- Selectively suppressing hippocampal sharp-wave ripples during post-training rest impaired rats' performance on a hippocampus-dependent spatial task, evidence that ripple-linked replay is causally involved in consolidation (Girardeau et al., 2009).
- Memories of everyday events depend at first on the medial temporal lobe and, as they mature, become increasingly dependent on neocortical regions such as prefrontal cortex (Frankland & Bontempi, 2005).
Human Measurement
Level of Measurement: Directly Measured
- The standard paradigm compares retention after an interval of sleep with retention after an equal interval awake; a meta-analysis of 823 effect sizes from 271 samples found a moderate sleep benefit for episodic memory (g = 0.44), reduced to g = 0.28 after correcting for selective reporting (Berres & Erdfelder, 2021).
- Across studies of total sleep deprivation after learning, memory was impaired with Hedges' g = 0.28 (95% CI 0.18 to 0.38), though the authors found publication bias and very low statistical power in most studies (Newbury et al., 2021).
- Targeted memory reactivation, replaying learning cues during sleep, improved later memory in a meta-analysis of 91 experiments (N = 2,004; g = 0.29, 95% CI 0.21 to 0.38), with effects in non-REM sleep but not in REM sleep or wakefulness (Hu et al., 2020).
- For motor sequence learning (34 articles, 1,296 participants), there was no evidence that sleep enhances learning once confounds such as time of testing and inhibition building up during training were controlled; sleep may stabilize skill memory, but that effect was inconsistent across designs (Pan & Rickard, 2015).
- Replication failures and boundary conditions suggest sleep's benefits for declarative memory are smaller, more task-dependent and less tied to slow-wave sleep than once assumed, so the human effect is real but modest (Cordi & Rasch, 2021).
AI Measurement
Level of Measurement: Indirectly Measured
AI Memory Performance: Not Yet Compared
- Widely used AI assistants update nothing in the model itself between conversations; a 2026 preprint tested nightly consolidation of conversations into the model's weights (its long-term learned parameters) and found it retained 80.4% of knowledge versus 36.8% for repeatedly summarizing the conversation, across 10 conversations and 1,146 questions (Dennis et al., 2026).
- Other 2026 preprints build explicit 'sleep' phases into research models, distilling recent in-context knowledge into long-term parameters with replay, or converting the recent conversation into persistent weights offline, and report better performance with longer sleep; these are prototypes scored against other AI methods, not tests against people (Behrouz et al., 2026; Lee et al., 2026).
- In training, facts a model picks up are not stabilized: each exposure raises their probability, but later training dilutes it through forgetting that follows a power law, measured as knowledge acquisition and forgetting rather than as consolidation (Chang et al., 2024).
- In saved memory, systems such as Mem0 extract and 'consolidate' facts from conversations into a store the AI can search later; this is scored only as part of broad long-conversation question-answering benchmarks, where Mem0 beat OpenAI's memory feature by 26% (relative) on a model-judged score (Chhikara et al., 2025).
- A 2026 preprint found that rewriting interactions into stored facts kept claims but erased the limits their source placed on using them (for example, a user's say-so becoming a standing instruction) in 48 of 49 combinations of seven memory systems and seven language models, so the failure was nearly uniform across systems; this resembles consolidation losing source detail but included no human comparison (Zhan et al., 2026).
- The closest classic paradigm tried on AI is interference: across 39 language models, earlier information in the conversation disrupted recall of later updates more than the reverse (d = 1.73), opposite to the typical human pattern, but this tests the context window (what the model can see in the current conversation) with no delay or offline period, and no people were tested (Chattaraj & Raj, 2026).
- Not yet compared: no study was found that gave AI and people the same consolidation test, such as retention after an offline period versus none; the AI results are gains of one engineered method over another, which cannot be placed on a common scale with the human sleep benefit of roughly g = 0.3 to 0.4 (Berres & Erdfelder, 2021).
Memory Processes · Acquisition and Transformation
Reconsolidation
Definition
- Reconsolidation is the finding that a consolidated long-term memory, once reactivated, can return to a temporarily unstable state and must be stabilized again to persist, so that post-retrieval plasticity mirrors the processes that first stabilized it after learning (Nader & Hardt, 2009).
- What separates it from ordinary consolidation is that the memory is already stable: disruption works only when the memory has just been retrieved, and the same treatment without reactivation leaves the memory intact (Nader et al., 2000).
- During the restabilization window the memory can be weakened, strengthened, or updated with new information, which distinguishes reconsolidation updating from extinction, where a new, competing memory is laid down alongside the old one (Lee et al., 2017).
- Retrieval alone is not sufficient: human fear-conditioning work shows that a prediction error, meaning a mismatch between what was expected and what happens at reactivation, is a necessary condition for the memory to destabilize (Sevenster et al., 2013).
Neural Substrate
- In rats, infusing the protein synthesis inhibitor anisomycin into the lateral and basal amygdala right after a fear memory was reactivated caused later amnesia, whether reactivation came 1 or 14 days after training; infusion without reactivation, or 6 hours after it, left the memory intact (Nader et al., 2000).
- Hippocampal memories also reconsolidate: anisomycin in the rat hippocampus erased a contextual fear memory only if it was reactivated first, even 45 days after training, and reactivation made a hippocampus-independent memory hippocampus-dependent again for about 2 days (Debiec et al., 2002).
- In humans, the beta-adrenergic (noradrenaline receptor) blocker propranolol given before reactivating a conditioned fear memory erased the fear-potentiated startle response 24 hours later and prevented the return of fear (Kindt et al., 2009).
- Human fMRI showed that reactivation followed by undisturbed reconsolidation left a memory trace in the basolateral amygdala that predicted later fear, whereas disrupting reconsolidation with extinction abolished that trace and weakened fear-circuit connectivity (Agren et al., 2012).
Human Measurement
Level of Measurement: Directly Measured
- Standard human paradigms reactivate a memory with a brief reminder and then intervene within the reconsolidation window (a drug, extinction training, or new learning), comparing against a group that receives the same intervention without reactivation (Lee et al., 2017).
- A meta-analysis of randomized double-blind studies found that propranolol given at reactivation reduced recall of aversive material and cue-elicited fear in 14 studies of healthy adults (n = 478; Hedges g = -0.51) and reduced symptoms in 12 clinical studies (n = 446; g = -0.42) (Pigeon et al., 2022).
- Across 63 comparisons, extinction delivered shortly after a reminder (post-retrieval extinction) reduced the return of fear in humans more than standard extinction, a small-to-moderate effect (g = 0.40) (Kredlow et al., 2016).
- For episodic memory, the reminder paradigm has people learn a list, receive a reminder or none the next day, and learn a second list; reminded participants later intrude second-list items into recall of the first list, and a meta-analysis confirmed that new learning after reactivation reliably increases such intrusions (Scully et al., 2017).
- The effect is not uniform: a high-powered, preregistered direct replication of the landmark reactivation-extinction study found no benefit over regular extinction, so the size and boundary conditions of human reconsolidation effects remain debated (Chalkia et al., 2020).
AI Measurement
Level of Measurement: Indirectly Measured
AI Memory Performance: Comparison Not Possible
- In training (the model's built-in knowledge, stored in its weights), knowledge-editing methods rewrite individual facts well but fail to carry the change through: on GPT-J, edited facts were recalled 88% (ROME) to 96% (MEMIT) of the time, yet accuracy on multi-hop questions that depend on them fell from 40.5% to about 7% (Zhong et al., 2023).
- Within the current conversation, when facts in the context window (what the model can see) conflict with what it learned in training, models accept coherent, convincing new evidence but also show confirmation bias toward their stored knowledge (Xie et al., 2024).
- A workshop paper adapted the human A-B, A-C proactive interference design to the conversation: as a value was repeatedly updated, models' accuracy at reporting the latest value declined log-linearly toward zero because they retrieved overwritten values; larger models held up better, but no human participants were tested (Wang & Sun, 2025).
- Saved-memory systems do update on retrieval in a loose sense: Mem0 retrieves stored memories similar to each new fact and has the language model choose to add, update, delete, or leave them, though this is an engineering design rather than a tested analogue of reconsolidation (Chhikara et al., 2025).
- A 2025 benchmark preprint tested whether agents overwrite older facts with newer contradicting ones: on single-fact questions scores ranged from 7% (Zep) to 78% (GPT-5-mini), and on multi-step questions every system scored 28% or lower, with most at 2-7%; no human baseline was reported (Hu et al., 2025).
- Comparison is not possible because there is no common scale: no study located runs AI on a reconsolidation paradigm, and human studies measure a change that depends on retrieval, against a no-reactivation control (e.g., Hedges g), whereas AI studies only score whether an update succeeded and report no human results (Pigeon et al., 2022).
Memory Processes · Acquisition and Transformation
Integrative Inference
Definition
- Integrative inference is the ability to form a relationship between items or events that were never experienced together, by linking separate memories that share an element (seeing a man with a dog, then the same dog with a woman, and inferring the two are a couple) (Zeithamova, Schlichting, & Preston, 2012).
- Two mechanisms are distinguished: integrative encoding, in which recalling a past event during a new overlapping experience links the new and old memories during learning, and retrieval-based recombination, in which separate memories are retrieved and combined only when a novel question is posed (Zeithamova, Schlichting, & Preston, 2012).
- The key distinction from ordinary associative memory is flexible expression: rats with hippocampal damage learned trained odor pairs normally but could not use them for transitive or symmetrical inferences about items never presented together (Bunsey & Eichenbaum, 1996).
- In humans, robust transitive inference goes with explicit awareness of the underlying hierarchy, marking it as a form of declarative memory: aware participants were near 100% correct on inference pairs while unaware participants performed poorly (Smith & Squire, 2005).
Neural Substrate
- Rats trained on four overlapping odor discriminations made correct transitive inferences, but after the hippocampus was disconnected from its cortical or subcortical pathways they still learned the separate problems yet showed no transitive inference (Dusek & Eichenbaum, 1997).
- In human fMRI, reinstating an earlier related event while encoding a new overlapping one engaged a functionally coupled hippocampus and ventromedial prefrontal circuit, and this activity tracked later success at inferring the untrained relationship (Zeithamova, Dominick, & Preston, 2012).
- Patients with focal ventromedial prefrontal damage (n = 5, versus 10 comparison participants) remembered directly studied AB and BC pairs normally but were significantly impaired at choosing the inferred AC pairs (Spalding et al., 2018).
- A meta-analysis of 32 fMRI studies found hippocampus, prefrontal cortex, posterior parietal and retrosplenial cortex engaged during transitive inference, with hippocampus, medial prefrontal and posterior cingulate cortex shared by hierarchical and associative inference tasks (Zhang et al., 2022).
Human Measurement
Level of Measurement: Directly Measured
- Two standard paradigms are built for this ability: associative inference (study overlapping pairs AB and BC, then test the never-studied AC pair) and transitive inference (learn premise pairs A>B, B>C, C>D, D>E, then test novel pairs such as B?D); enough studies use them that a 2022 meta-analysis could pool 32 fMRI experiments (Zhang et al., 2022).
- In a transitive inference study of 56 adults, inference was at chance (52%) 20 minutes after learning but exceeded 75% after 12 or 24 hours, and for the most distant pairs was 93% after sleep versus 69% after an equal period awake (Ellenbogen et al., 2007).
- Across four associative inference experiments (24 young adults each, tested 1 to 2 days after study), mean accuracy on inferred AC pairs ranged from 64% to 80%, closely matching memory for the directly studied pairs (66% to 78%) (Carpenter & Schacter, 2017).
- Individual spread is large on transitive inference: participants aware of the hierarchy scored near 100% on inference pairs, while unaware participants and amnesic patients with hippocampal damage performed poorly on the critical BD pair (Smith & Squire, 2005).
- In an associative inference task given to 83 young and 80 older adults, older adults showed a deficit in inference that was larger than their deficit in direct associative memory, showing the paradigm separates inference from memory for the pairs themselves (Charles & Bowman, 2025).
AI Measurement
Level of Measurement: Directly Measured
AI Memory Performance: Not Yet Compared
- Transitive inference has been tested by name in large language models given the premises in the conversation (in the model's context window, meaning what it can see at the moment): GPT-3.5-Turbo, GPT-4, Llama3-8B and Qwen showed strong inference on 10-item hierarchies with human-like symbolic distance and terminal-item effects, but the abstract reports no human group and no accuracy figures (Wu & Deng, 2025).
- The systems differed in how they responded to help: step-by-step (chain-of-thought) and confidence prompts improved GPT-4 on transitive inference but lowered accuracy in the other models tested, and terminal-item biases persisted or grew (Wu & Deng, 2025).
- A preprint gave small transformers the classic animal and human training procedure (adjacent pairs only): when pairs were learned into the model's weights (its long-term, trained-in knowledge) the models inferred non-adjacent pairs above chance with a symbolic distance effect, but when pairs were only shown in context they showed no transitive inference unless specially pre-trained; in all but the smallest of seven Gemini and Gemma models a 'number line' prompt improved inference over a circle prompt (Geerts et al., 2025).
- For facts learned in training, a preprint fine-tuned Llama 3 8B, Qwen 2.5 7B, GPT-4o-mini and GPT-4o on invented facts placed in separate documents: all models answered single facts accurately but were at chance on questions linking two facts unless allowed to reason step by step in writing, when accuracy became high; in Llama 3 8B, seeing both facts together in one document or in the conversation gave non-zero accuracy (Balesni, Korbak, & Evans, 2024).
- Using real facts from pretraining and excluding questions whose answer could have been memorized directly, the best models linked two facts without writing out intermediate steps 80% of the time when the linking item was a country but only 5% of the time when it was a year, a wide spread across fact types (Yang et al., 2025).
- For saved memory, a long-conversation benchmark asked questions that require combining information from different sessions: across five systems multi-hop scores ranged from 12.8 to 42.0 F1 (GPT-3.5-turbo-16K best) against a listed human score of 85.8, but the paper does not report how many people answered or whether they could reread the conversations (Maharana et al., 2024).
- Not yet compared: no study gives people and AI the same integrative inference test, though one could; the only human score beside AI (Maharana et al., 2024) lacks sample and procedure; in-context tests let AI see the premises while people must remember them, so strong AI results there cannot count; and training-based failures come from a no-written-reasoning condition people are not held to, with no human data on a shared scale (Balesni, Korbak, & Evans, 2024).
Memory Processes · Retrieval and Surfacing
Recognition (Familiarity and Recollection)
Definition
- Recognition is the judgment that something has occurred before, and it can rest on two bases: a context-free sense that the item is familiar, or retrieval of the context in which it was met, as when a man on the bus seems familiar before you place him as the butcher from the supermarket (Mandler, 1980).
- Unlike recall, recognition presents the item itself as the cue; within recognition, recollection retrieves specific details of the original episode, while familiarity is a graded feeling of prior occurrence without those details (Yonelinas, 2002).
- The remember/know distinction separates recognition accompanied by conscious re-experiencing of the study episode (remember) from recognition based on knowing an item was seen without recalling the occasion (know) (Tulving, 1985).
- Whether familiarity and recollection are truly separate processes is still debated: a single continuous memory-strength account (unequal-variance signal detection) fits recognition confidence data as well, and individual recognition decisions are unlikely to be process-pure (Wixted, 2007).
Neural Substrate
- Reviewing lesion, imaging and single-cell work in humans, monkeys and rats, the hippocampus appears critical for recollection but not familiarity, parahippocampal cortex contributes contextual information to recollection, and perirhinal cortex supports familiarity-based recognition (Eichenbaum et al., 2007).
- Patient NB, whose surgical resection removed much of left perirhinal cortex but spared the hippocampus, showed impaired familiarity with preserved recollection, the dissociation that separate-process accounts predict (Bowles et al., 2007).
- The clean split is contested: five patients with bilateral lesions thought limited to the hippocampus showed a broad recognition deficit that reduced both the recollection and familiarity parameters of a dual-process model, and both parameters of a single-process model (Dede et al., 2013).
- A meta-analysis of 41 EEG experiments (about 1,000 participants) found the dissociation dual-process models predict between a mid-frontal brain response at roughly 300–500 ms, linked more to familiarity, and a parietal response at roughly 500–800 ms, linked more to recollection (Kwon et al., 2023).
Human Measurement
Level of Measurement: Directly Measured
- Recognition is tested with purpose-built old/new and two-alternative forced-choice tests, and familiarity and recollection are separated with remember/know judgments, confidence-based receiver operating characteristics (ROCs) and process dissociation; a meta-analysis of 25 reports using these methods found healthy aging reduces recollection (d = −0.75, 95% CI −0.85 to −0.65) far more than familiarity (d = −0.27, 95% CI −0.45 to −0.09) (Koen & Yonelinas, 2014).
- In the classic forced-choice study, people looked once through about 600 randomly chosen words, sentences or pictures and then picked the old item from pairs with a randomly chosen new one; the median participant was correct on 90% of word pairs, 88% of sentence pairs and 98% of picture pairs (Shepard, 1967).
- A contemporary summary gives the sentence procedure in more detail: college-age students inspected 612 short, unrelated sentences one by one and then chose the studied member of 68 test pairs, averaging 89% correct (50% is chance) (Carroll, 1971).
- Fourteen adults who viewed 2,500 object pictures over 5.5 hours then picked the studied object at 92.5% (SEM 1.6) against a new category, 87.6% (SEM 1.8) against a similar exemplar, and 87.2% (SEM 1.8) against the same object in a different state (Brady et al., 2008).
- Forced-choice scores do not isolate recollection: a patient with selective hippocampal damage was unimpaired at forced-choice object recognition yet impaired at recall and at yes/no recognition with very similar foils, consistent with forced-choice tests being passable on familiarity alone (Holdstock et al., 2002).
AI Measurement
Level of Measurement: Indirectly Measured
AI Memory Performance: Not Yet Compared
- A 2023 preprint gave 19 language models (GPT-2, 2019; GPT-Neo and GPT-J, 2021; OPT and BLOOM, 2022; 125 million to 13 billion parameters) a sentence test modeled on Shepard's: each model was trained once on 600 new sentences, then shown each studied sentence with a new one, but it made no old/new choice; the author counted a trial correct when the model's loss (how surprising it found the text) was lower for the studied sentence, explicitly as "a proxy for familiarity" (Orhan, 2023).
- After one exposure, every model's loss comparison picked the studied sentence on roughly 94% (smallest GPT-2) to nearly 100% (best BLOOM, OPT and GPT-J models) of pairs, plotted against a human band of about 88–91%; the human value is not stated in the text, no human sample or spread is given, and no statistical test against people was run (Orhan, 2023).
- The human baseline is only partly confirmed: Shepard's abstract gives a median of 88% for sentences and a 1971 summary gives 612 sentences, 68 test pairs and 89% correct, which fits Orhan's line, but the number of participants and the spread across them could not be confirmed because the original article was not accessible (Shepard, 1967).
- The design matches Shepard's in its essentials (one exposure, randomly chosen new foils, a two-choice test straight after study), but the scoring does not: people made a recognition decision, while the models' familiarity-like strength signal was read off by the experimenter, learning settings were searched to maximize memory (about 20,000 runs across the study), models were scored on 600 pairs rather than 68, and the sentences differed; a result favouring the side these differences helped cannot be counted (Orhan, 2023).
- No study located tests the recollection route in AI (remember/know judgments, an ROC recollection estimate, or retrieving the study context of a recognized item); the nearest is a 2024 study that trained a language model on synthetic documents to cite the document a fact came from, a source-memory task with no human baseline, so a familiarity-only AI signal is being set against a human score that can draw on both routes (Khalifa et al., 2024).
- Recognition of text met once in full-scale pretraining is far weaker: across Pythia models (2023) of 160 million to 12 billion parameters trained on the Pile, familiarity-based detectors of whether a passage was in the training data performed near chance (AUC below 0.6, where 0.5 is chance) in most domains; no human baseline was run (Duan et al., 2024).
- When chat models made an explicit choice of the verbatim passage from a likely-trained book among four options (the original plus three paraphrases, 25% chance), ChatGPT scored 72.0% and Claude 2.1 (2023) 73.4%; the 10 human raters had not studied the books and scored near chance, so this checks the task, not human memory (Duarte et al., 2024).
- In the current conversation, an item recognition test run on Pythia models (14 million to 12 billion parameters) and GPT-2, scored from how strongly the model predicted list items rather than from yes/no answers, found strong primacy but weak or absent recency, unlike people; the list stayed visible to the model and no human data were collected (Clark & Schuler, 2026).
- For saved memory, benchmarks test only something close: LongMemEval includes 30 false-premise questions requiring a chat assistant to recognize that information never appeared in past sessions, with no human baseline; with recognition measured in training and the conversation but not saved memory, and familiarity but not recollection, the AI level is indirect and no comparison with documented human results has yet been made (Wu et al., 2025).
Memory Processes · Retrieval and Surfacing
Temporal Placement
Definition
- Memory for when events happened covers their order, how long ago they occurred and their date. Friedman's review sorted the theories into three kinds: an impression of distance (such as how strong or accessible a memory feels), reconstruction of a location in time from remembered context and known time patterns, and stored before-and-after relations, with reconstruction the best supported (Friedman, 1993).
- Episodic time is the tracking and organizing of events in time as they are experienced, stored and retrieved from memory, covering both specific moments and the durations between them (Kwok et al., 2025).
- Relative temporal memory (which event came first, or which phase an item belonged to) is separable from absolute temporal memory (how far apart or how long ago): a meta-analysis found emotion improves the first but not the second (Ye & Ke, 2026).
- Placing an item in time ("when did you see this image?") is a different judgment from recognizing it ("did you see this image?"). The two diverge when items linked to a frightening episode are well recognized but wrongly assigned to the phase in which that episode happened (Laing et al., 2026).
Neural Substrate
- Patients with frontal lobe lesions were impaired at putting 15 words, or 15 public events from 1941 to 1985, into chronological order despite normal recall and recognition of the items; patients with Korsakoff's syndrome showed an order deficit out of proportion to their amnesia (Shimamura et al., 1990).
- In freely foraging rats, the population activity of the lateral entorhinal cortex encoded time on scales from seconds to hours, a signal not found to the same degree in medial entorhinal cortex or hippocampal areas CA3 and CA1 (Tsao et al., 2018).
- In human fMRI, the similarity of hippocampal activity patterns across items at encoding predicted whether people later judged those items to have occurred close together in time when the context had changed between them (Ezzyat & Davachi, 2014).
- Hippocampal pattern similarity across the items that came between two pictures predicted later memory for their order, and the content of those in-between items was reactivated while people made recency judgments (DuBrow & Davachi, 2014).
Human Measurement
Level of Measurement: Directly Measured
- The standard tests are temporal order and recency judgments (which of two items came first), temporal source judgments (which phase an item came from) and temporal distance judgments. A three-level meta-analysis pooled 33 studies (44 experiments, 79 effect sizes) of these tests, finding emotion improved order memory (g = 0.35, 95% CI 0.11 to 0.59) and source memory (g = 0.28) but not distance judgments (g = 0.06) (Ye & Ke, 2026).
- A meta-analysis of 90 studies (3,197 younger and 3,192 older adults) found that memory for temporal order, like memory for source and spatial location, declines more with age than memory for the items themselves (Old & Naveh-Benjamin, 2008).
- When 1,579 adults dated news events online, very recent events were placed almost exactly (median displacement 1 day), events 100 to 1,000 days old were pushed back in time (median 23 days, backward telescoping) and older events were pulled forward (median 365 days, forward telescoping); calendar dates were more accurate than "how long ago" estimates (Janssen et al., 2006).
- Psychologists dating colloquia they had attended made errors that grew by about 0.4 days for every day since the talk, and the errors were unbiased in direction, a pattern that on its own can produce telescoping in surveys (Rubin & Baddeley, 1989).
- In a book-length order test (a 2024 preprint), 155 readers of The Murder of Roger Ackroyd, tested on average 7.5 days after finishing it, judged which of two passages came first (10 pairs each). Accuracy averaged 64% for 50-word and 56% for 20-word passages, reached 76% for passages more than 25,000 words apart, and was reliably above chance (50%) only beyond 4,000 words (Pink et al., 2024).
AI Measurement
Level of Measurement: Directly Measured
AI Memory Performance: Below Human approximate provisional
- A 2024 preprint adapted recency-judgment tasks from cognitive psychology into a passage-order test for language models, using books unlikely to be in their training, and tested all three places AI holds information. Although framed as episodic memory, the task is a relative-order (recency) judgment, so it bears on this mechanism directly; it does not test dating or how-long-ago judgments. Before seeing the books, 11 models scored 49% to 57% (GPT-4 highest) (Pink et al., 2024).
- Deciding comparison, the one where neither side could look at the book: Llama 3 8B and Mistral 7B (2023–2024 releases) further trained on the books and then tested without them scored about 50% to 51% across five rounds of training, no better than copies trained without the books (McNemar tests, p > .05), while readers averaged 56% to 64% depending on passage length. The tests differ (nine books versus one, repeated training versus one reading, different passage pairs), but the result is clear because the models stayed at chance while readers were well above it (Pink et al., 2024).
- An ICML 2026 paper gave the same 296 passage pairs to 97 readers (tested within 30 days of finishing the novel) and to models holding the whole novel in their context window, meaning the text they can see in the current conversation. Both showed the human distance effect; Llama 3.1 8B did not differ from readers at any distance (all p ≥ .289), and Llama 3.1 70B (both 2024 releases) beat them only for passages 16,000–24,000 words apart (about 92% versus 72%) (Pink et al., 2026).
- That match was on a test easier for the AI (the models could still see the book while readers had to remember it), so it counts only as a provisional Comparable and does not outrank the training result, which tests memory rather than search over visible text. Across systems the spread is wide: with short excerpts visible, 10 models ranged from 56% (Llama 2 7B) to 95% (Mixtral 8x22B); with retrieval-augmented generation, meaning a system that searches the stored book and pastes matching passages into the prompt, accuracy was 55–67% (Pink et al., 2024).
- For saved memory in chat assistants, a benchmark of questions about past conversations found that ChatGPT's memory feature answered 44% (GPT-4o) to 65% (GPT-4o-mini) of temporal questions and Coze 4% to 39%, and that tagging stored items with event dates improved retrieval for temporal questions by about 7% to 11%. The "human evaluation" in that study means people scored the systems; no human group took the test (Wu et al., 2025).
- For what a model learned in training, the dates of historical figures' deaths, works of entertainment and news headlines could be read out of Llama 2's internal activity with R² between 0.56 and 0.89, rising with model size. No human dating data were collected on the same items, so there is no common scale with human event dating and its telescoping errors (Gurnee & Tegmark, 2024).
- Models also misplace their own knowledge in time: Llama 2, trained on text up to 2022, mostly answered time-sensitive questions with facts as of about 2019 until it was tuned toward a target year. This resembles misdating one's own memories, but no human group was tested (Zhao et al., 2024).
Memory Processes · Retrieval and Surfacing
Context-Dependent Retrieval
Definition
- The encoding specificity principle holds that a retrieval cue helps only to the extent that its information was stored together with the target at learning, so what counts as a matching context is fixed by how the event was encoded, not by how strongly the cue and target are related in general (Tulving & Thomson, 1973).
- State-dependent memory is the internal-state version: men who learned material while intoxicated recalled it better 24 hours later when intoxicated again than when sober, while recognition was not significantly affected by the change of state (Goodwin et al., 1969).
- What separates context dependence from ordinary cueing is that state and context effects appear when retrieval happens without observable reminders, as in free recall, and typically vanish when discrete retrieval cues are supplied at test (Eich, 1980).
- Incidental environmental context (the room, the setting) is distinguished from focal cues: its effects shrink when non-contextual cues dominate at learning (overshadowing) or at test (outshining), or when people mentally reinstate the original context (Smith & Vela, 2001).
Neural Substrate
- In fMRI during free recall, brain activity patterns specific to the category of pictures studied (faces, locations, objects) reappeared several seconds before people recalled items from that category, consistent with reinstated study context cueing the memory search (Polyn et al., 2005).
- Intracranial recordings from 69 neurosurgical patients showed that recalling a word re-evoked activity patterns from the study of neighbouring list items, falling off with list distance and strongest in the temporal lobe; patients with a stronger signal more often recalled neighbouring items in succession (Manning et al., 2011).
- Activity patterns for individual word-scene pairings were reinstated in parahippocampal cortex when the scene was recollected from the word alone, and the degree of reinstatement tracked hippocampal activity, pointing to the hippocampus as coordinator of cortical pattern completion (Staresina et al., 2012).
- Rodent and human studies of fear conditioning and extinction implicate a hippocampus-amygdala-medial prefrontal circuit in making learned responses depend on the context in which they were acquired or extinguished (Maren et al., 2013).
Human Measurement
Level of Measurement: Directly Measured
- In the classic context-match paradigm, divers learned word lists on land or underwater and recalled them in the same or the other environment; lists were best recalled where they were learned, an interaction (F = 22.0) that a later reanalysis converts to a lower-bound Cohen's d of about 2.7 (Godden & Baddeley, 1975).
- A replication with 16 divers, each tested in all four learn-recall combinations, found no same-environment advantage (interaction F(1, 60) = 2.0, p = .16); the author notes that environmental context effects in free recall are usually modest, around d = 0.25 (Murre, 2021).
- A meta-analysis of incidental environmental context studies found context effects reliable overall, but reduced by overshadowing, outshining and mental reinstatement of the learning context at test (Smith & Vela, 2001).
- A 2026 meta-analysis of 33 studies (106 effect sizes, N = 2,177) found that reinstating the learning context reliably improved memory (Hedges' g = 0.32), equally in younger and older adults, with weaker effects for associatively encoded word pairs and familiar switched contexts (Symeonidou et al., 2026).
- The Cognitive Interview, whose techniques include mentally reinstating the context of the witnessed event, was meta-analysed across 46 articles: it produced a large, significant increase in correct details with only a small increase in errors (Memon et al., 2010).
AI Measurement
Level of Measurement: Indirectly Measured
AI Memory Performance: Not Yet Compared
- For what a model learned in training, cue dependence has been studied under another name, the reversal curse: a preprint found GPT-4 answered celebrity-parent questions correctly 79% of the time when cued in the direction the fact is usually written, but 33% when cued in reverse, whereas the same facts presented in the conversation could be reversed (Berglund et al., 2023).
- Within the context window (what a model can see in the current conversation), retrieval suffers when the question shares few words with the stored fact: at 32,000 tokens, 11 of 13 long-context models fell below half their short-text score, and even the strongest, GPT-4o, dropped from 99.3% to 69.7%; this mirrors cue-trace overlap but has no human baseline (Modarressi et al., 2025).
- A 2024 preprint gave one model (Mistral-7B) Tulving and Watkins's cue-valence experiment, built on encoding specificity, with the word list still visible in the chat: associative cues recalled targets about as often as in the original human data (0.49 and 0.38 vs 0.52 and 0.43), while rhyme cues did far worse (0.11 and 0.14 vs 0.45 and 0.43) (Chauvet, 2024b).
- Not yet compared: an AI has no physical surroundings, but the context of a conversation can be matched or changed between learning and test, and Tulving's paradigms have been run on a model; these same-paradigm results still cannot be scored under the rules: the preprints (this one also ran Tulving's recognition-versus-recall test, with copy-cue recall 0.46 vs 0.69) give only human proportions from the 1970s-80s, with no sample size or statistical test, test one small model, and the rhyme gap partly reflects the model's weak rhyme knowledge rather than context matching (Chauvet, 2024a).
- Induction heads (attention components that find an earlier occurrence of the current word and retrieve what followed it) behave like the temporal-context model of human free recall across five model families from GPT-2 to Llama-3-8B, with fitted context-drift parameters in the range of human data (N = 171); this compares the shape of a recall bias on still-visible text, not retrieval success (Li et al., 2024).
- A 2026 preprint gave Llama 3.1 8B and 70B a whole novel in context and 97 readers the same 296 which-came-first judgments from memory; both showed the human distance effect, 8B did not differ from people and 70B beat them at one distance range, and one 'time-reinstatement' attention head drove performance (ablation cost 10-15%), but it tests memory for the order of events rather than retrieval helped by a matching context, so it is not used for this rating (Pink et al., 2026).
- For saved memory, a 2026 preprint that stored each fact in an AI agent's long-term memory together with a narrative of the situation in which it was learned raised cross-session recall on the LongMemEval benchmark from 53.5% to 73.7%, an engineering use of encoding specificity with no human comparison (Stern & Nadel, 2026).
Memory Processes · Retrieval and Surfacing
Involuntary Retrieval
Definition
- Involuntary autobiographical memories are memories of personal events that come to mind spontaneously, with no conscious initiation of the retrieval process; they arise through associative processes with minimal executive control and favour recent events that overlap with features of the current situation (Berntsen, 2021).
- The key contrast is with voluntary memories, which are retrieved strategically; in a one-day diary study of 48 undergraduates, involuntary memories were reported three times as often as voluntary ones and were more linked to daydreaming and unfocused attention (Rasmussen & Berntsen, 2011).
- Involuntary memories are distinguished from mind wandering by their automatic, situation-dependent triggering by cues, and they occur in young children and apes, suggesting an early-emerging, evolutionarily old retrieval mode (Berntsen, 2021).
- Intrusive traumatic memories are a clinical subtype rather than the whole construct: a review of 79 studies found involuntary memories are more frequent across many mental disorders and especially negative and distressing in PTSD and depression, marking them as a transdiagnostic phenomenon (Shan et al., 2025).
Neural Substrate
- In fMRI, sound cues that spontaneously evoked paired scenes engaged the same medial temporal, ventral occipitotemporal and ventral parietal retrieval network as deliberate recall, while deliberate recall recruited more dorsal frontal activity (Hall et al., 2014).
- Combined fMRI and ERP work found that incidental retrieval, like intentional retrieval, involves the hippocampus, precuneus and ventrolateral prefrontal cortex when successful, but is initiated by early bottom-up signals in posterior cortex (around 170 ms after the cue) rather than dorsolateral prefrontal control (Kompus et al., 2011).
- In 35 adults scanned while viewing traumatic film footage, scenes that later returned as intrusive memories showed heightened activity at encoding, and the involuntary recall itself was marked by frontal activity, notably the left inferior frontal gyrus (Clark et al., 2016).
- Compared with 21 trauma-exposed controls, 21 people with PTSD reported more, and more vivid and emotional, involuntary memories that activated a similar network of regions but with delayed neural responses (Hall et al., 2018).
Human Measurement
Level of Measurement: Directly Measured
- The standard laboratory test is the vigilance task: people watch for rare vertical lines among horizontal ones while irrelevant cue phrases appear, and write down any memories that pop up; these memories were more specific and retrieved faster than word-cued voluntary memories, and did not differ from diary-recorded everyday involuntary memories in the same people (Schlagman & Kvavilashvili, 2008).
- The vigilance paradigm has become an often-employed method reused by independent labs, for example showing that an added demanding task lowers the number of involuntary memories reported across two studies (Barzykowski & Niedźwieńska, 2018).
- A second paradigm pairs unique or repeated sounds with scenes at encoding and later presents the sounds during an unrelated task; unique sounds produced the most involuntary memories, which were processed faster than voluntary ones (Berntsen et al., 2013).
- Across a film-based laboratory study and a daily-life study, older adults reported involuntary episodic memories as often as younger adults, even though their voluntary remembering was slower or less frequent (Berntsen et al., 2017).
- Intrusive memories, counted in diaries after a distressing film (the trauma film paradigm), have been pooled in meta-analysis: across 134 articles, 606 effect sizes and 12,074 non-clinical participants, interventions reduced intrusion frequency by a small amount (g = 0.16, 95% CI 0.09 to 0.23) (Varma et al., 2024).
AI Measurement
Level of Measurement: Indirectly Measured
AI Memory Performance: Comparison Not Possible
- No study was located that tests AI on involuntary retrieval by name or runs the vigilance-task, sound-cue or diary paradigms on an AI system; the closest evidence comes from studies of unrequested recall under other names (training-data reproduction, priming, proactive preference use, memory leakage), so the level is indirect (Aerni et al., 2025).
- From training: when given ordinary requests such as writing a letter or tutorial, chat models reproduced text they had learned from the internet without being asked to, with 50-plus-character matching snippets making up about 7% of output for Gemini 1.5 up to more than 15% for Claude 3 Opus across four model families (Aerni et al., 2025).
- From training: after a new fact is learned through training updates, models show a 'priming' effect in which the new keyword intrudes into unrelated contexts after as few as 3 spaced exposures, most strongly for words the model found improbable beforehand, across PaLM-2, Gemma and Llama models (Sun et al., 2025).
- In the current conversation (the model's context window, meaning what it can see right now): on PrefEval, models had to bring up a preference stated earlier when a later request made it relevant, without being reminded; on one topic, zero-shot accuracy was 0.50 for the best model (o1-preview) at 10 turns, falling to 0.14 at 300 turns, versus 0.07 and 0.02 for Claude 3.5 Sonnet, and most of 10 models were below 10% at 10 turns (Zhao et al., 2025).
- In saved memory (stored user facts carried into later tasks): models surfaced stored personal attributes in tasks where they did not belong in up to 69% of cases, and GPT-5's rate rose from 0.1% to 9.6% as tasks increased from 1 to 40, a rough counterpart of unwanted cue-triggered recall rather than a test of it (Mireshghallah et al., 2026).
- Memory-equipped agent designs already contain a part that does this job: the generative-agents architecture stores a running record of experiences and retrieves relevant ones automatically at each step to guide behaviour, without an explicit request to remember, but that retrieval was evaluated for believability, not against human involuntary memory (Park et al., 2023).
- Comparison is not possible because there is no common scale with documented human results: the only human baseline found (median zero overlap for human writing versus up to 7.5% median for models) used online posts with no count of people, no controlled procedure, possible internet access and some plagiarism, and it measures verbatim copying rather than memories coming to mind; none of the other AI studies included people (Aerni et al., 2025).
Memory Processes · Retrieval and Surfacing
Standing Constraints
Definition
- The multiprocess framework describes two routes by which stored memories guide later behaviour: effortful monitoring, and spontaneous retrieval, in which something processed as part of the ongoing activity reflexively brings the relevant memory to mind; the framework was built for intentions, but spontaneous retrieval is the route a standing constraint needs, since nothing in the moment names the fact (McDaniel & Einstein, 2000).
- Research on audience design, the tailoring of speech to a particular listener, treats the use of partner-specific knowledge as a by-product of ordinary memory: speakers adapt to a partner when the partner and the relevant information are readily associated in memory, rather than through a special search for what that person knows (Horton & Gerrig, 2005).
- Knowing a fact about another person does not guarantee using it: adults who clearly knew a partner could not know what was hidden in a bag still often took the partner's instructions to refer to the hidden object, a dissociation between reflective knowledge and its routine use (Keysar et al., 2003).
- The general failure is called knowledge neglect, not relying on stored knowledge that is available; standing constraints are a person- or thing-specific case, distinct from the general-knowledge cases usually studied, such as rating a repeated false statement as true despite knowing better (Fazio et al., 2015).
Neural Substrate
- An activation likelihood meta-analysis of prospective memory imaging found that cues central to the ongoing task, which favour spontaneous retrieval, engage ventral parietal cortex, posterior cingulate and cerebellum, whereas cues that require active watching engage left anterior prefrontal cortex (BA 10) and anterior cingulate (Cona et al., 2016).
- An fMRI study found two dissociable routes to successful prospective remembering: sustained activity in anterior prefrontal attentional-control regions for one task, and purely transient activity in parietal and ventral regions linked to attentional capture and episodic retrieval for the other (McDaniel et al., 2013).
- Five patients with hippocampal amnesia used information shared with a conversation partner normally when it had just been mentioned, but failed markedly after a filled delay of about 40 seconds, compared with five matched comparison participants, implicating declarative memory in keeping partner-specific knowledge available (Rubin et al., 2011).
- In three-party conversations, people with bilateral hippocampal damage showed weakened tailoring to which partner already knew which images, and none in one experiment, while matched comparison participants tracked it; some partner-specific adaptation survives without the hippocampus, but flexible use depends on it (Yoon et al., 2024).
Human Measurement
Level of Measurement: Indirectly Measured
- No standard experiment isolates whether a known fact about a person surfaces unprompted and shapes a decision; the nearest work is on common ground in conversation, and a systematic review of 23 matching-task studies found seven different linguistic measures in use, with no single agreed test (Bovet et al., 2024).
- In a representative audience-design study, neurotypical adults (4 and 9 people in two experiments) gave reliably longer descriptions to a partner who had not seen an image than to one who had, and later identified which partner knew which image with 93.75% accuracy (Experiment 2); such results are reported as regression effects, not as rates of using versus missing a fact (Yoon et al., 2024).
- Event-based prospective memory, in which a cue appears unannounced inside another task, is the closest replicated paradigm: a meta-analysis of 117 effect sizes from 4,709 participants confirmed that cues central to the ongoing task are handled more easily, but the paradigm tests remembering to carry out an intention, not applying a fact about someone (Kliegel et al., 2008).
- The largest body of data on decisions shaped by another person's known preferences comes from surrogate decision-making: across 16 studies and 2,595 surrogate-patient pairs, surrogates predicted patients' treatment preferences with 68% accuracy, and prior discussion did not improve accuracy, but surrogates were asked directly, so the fact did not have to surface on its own (Shalowitz et al., 2006).
- Everyday records show standing constraints failing: among 512 young children with milk or egg allergy followed for about three years, 52.5% had more than one reaction, people other than parents supplied the food in 50.6% of reactions, and 11.2% of reactions to milk, egg or peanut followed purposeful exposure, though such records cannot separate memory lapses from label-reading errors or cross-contact (Fleischer et al., 2012).
AI Measurement
Level of Measurement: Indirectly Measured
AI Memory Performance: Below Human provisional
- A human comparison does exist, on the same items: ImplicitMemBench gave five people and seventeen AI systems the same 300 tasks under the same protocol, and the people answered all of them correctly while the best system, DeepSeek-R1, reached 65.3%, with Qwen3-32B at 64.1% and GPT-5 at 63.0%; the split by kind of constraint is the sharper result, 75.0% when a stored fact supports the answer against 17.6% when it rules something out (Qin et al., 2026).
- Current conversation (what the model can see in its context window): in PrefEval, where a user states a preference, often an avoidance such as an allergy, phobia or refusal, and later asks for advice that a generic answer would violate, most of 10 models followed it less than 10% of the time after 10 turns; the best, o1-preview, managed 50% at 10 turns and 14% at 300 turns, yet a one-line reminder raised it to 98%, so the fact was still visible but did not surface on its own (Zhao et al., 2025).
- Facts that rule something out are hardest: in a 2025 preprint, when an earlier remark (such as a healing leg) should override a known liking, answer accuracy with GPT-4o-mini was 1.6% to 7.4% across memory methods, against 60% to 79% when the earlier fact supported the answer, and the strongest model tried, o1, reached at best 29.7%; PrefEval, often recalled for this, did not itself compare likes with dislikes (Li et al., 2025).
- Safety-critical facts within one short conversation: a 2024 preprint that disclosed a severe allergy, phobia or physical limitation at the start found overall pass rates of 88.4% (Llama 3.1 405B) and 85.5% (o1-preview), while GPT-3.5 Turbo scored 27.9% even in the simplest scenario; adding other people's conflicting wishes cut the average from 75.1% to 43.2%, and telling models to consider the user's constraints raised one tested scenario to 94.6% (Alberts et al., 2024).
- Saved memory and long histories (a 2026 preprint testing 16 systems on 1,000 preferences): systems that could state a preference when asked often failed to use it in a later scenario that did not mention it; GPT-4o-mini recalled explicit preferences 83.7% of the time but acted on them 19.8% of the time, and average act accuracy ranged from 11.2% (Zep/Graphiti memory) to 61.5% (Claude 4.6 Sonnet), with health-related preferences used least by memory systems (Feng et al., 2026).
- A 2026 preprint framed the ability as prospective memory, testing it under that name in the conversation and in saved-memory add-ons: GPT-5.2 with high reasoning scored 83.2% overall (from 40.7% for Qwen3-32B) but was correct on only 43.9% of control items where the risk had already been resolved, mostly by giving needless reminders, memory add-ons (Mem0, A-MEM, Letta) scored 41.4% to 62.3%, and GPT-4o answered direct questions about the same facts at 98% or more while spontaneous use fell below 40% in long contexts (Zhang et al., 2026).
- The benchmark most like LoCoMo that tests this ability is the 2026 LoCoMo-Plus preprint, which adds earlier constraints that are never asked about: the best model, Gemini 2.5 Pro, respected them 26.1% of the time against 71.8% on LoCoMo's ordinary recall questions, and add-on memory systems scored 14.9% to 17.2%; LoCoMo itself asks only direct questions and has no preference or dislike condition (Li et al., 2026).
- Only one test has documented human results, and it is a thin baseline: ImplicitMemBench's five participants were all doctoral students and all scored 100%, which shows the tasks are solvable rather than establishing how well people do them, so the direction of the gap is firmer than its size; of the others, none tested people, LoCoMo's human score (F1 87.9) came from in-house annotators with no sample reported, on direct questions, and the Moses-illusion comparison in training knowledge tests noticing a false premise against general knowledge, a different ability (Qin et al., 2026).
Memory Processes · Retrieval and Surfacing
Prospective Initiation
Definition
- Prospective memory is becoming aware of a previously formed plan at the right time and place; it is distinguished from vigilance (watching continuously for a target held in mind) and from habitual routines, and from retrospective memory, which is recalling past content when asked (Uttl, 2008).
- Event-based prospective memory is triggered by an external cue (acting when a target word or person appears), whereas time-based prospective memory must be self-initiated at a set time or after a set interval, with no external cue to prompt it (Einstein & McDaniel, 1990).
- The multiprocess view holds that the intention can be retrieved either by effortful strategic monitoring of the environment or by spontaneous retrieval triggered when the cue appears, and people rely on each to different degrees depending on task demands (Einstein et al., 2005).
Neural Substrate
- In a voxel-based lesion study of 45 patients with focal brain lesions and 107 controls, damage to right rostral (polar) prefrontal cortex in Brodmann area 10 was specifically linked to deficits in time-based prospective memory, not explained by basic attention or inhibition problems (Volle et al., 2011).
- Imaging studies show a medial-lateral split in rostral prefrontal cortex: lateral regions are more active while a delayed intention is held during an ongoing task, and medial regions more active when only the ongoing task is performed (Burgess et al., 2011).
- A neuroimaging meta-analysis found that holding an intention recruits the dorsal frontoparietal (top-down attention) network, while retrieving it when the cue arrives recruits the ventral frontoparietal (bottom-up attention) network, with anterior prefrontal cortex involved across phases (Cona et al., 2015).
- A 7-tesla fMRI study of 58 healthy adults aged 60 to 75 found that event-based tasks engaged perceptual and attentional regions while time-based tasks engaged cognitive-control and buffering regions, and incentives improved only event-based performance (Peter et al., 2026).
Human Measurement
Level of Measurement: Directly Measured
- The standard laboratory paradigm embeds a rarely occurring prospective-memory cue in an ongoing task and scores whether the participant makes the intended response without being reminded; a meta-analysis of 117 effect sizes from 4,709 participants pooled results from this paradigm (Kliegel et al., 2008).
- Age effects are reliable in the lab: younger adults outperformed older adults on time-based (r = -.39) and event-based (r = -.34) tasks, yet older adults did better in naturalistic everyday tasks (r = .35 event-based, r = .52 time-based) (Henry et al., 2004).
- Age effects are larger when the cue is nonfocal (not part of what the ongoing task already requires attending to) than when it is focal, though older adults are impaired with both cue types (Kliegel et al., 2008).
- Holding an intention slows the ongoing task (a lexical decision task), and participants who slowed more caught more cues, a cost not seen when the embedded task was a retrospective memory task, showing that initiating a delayed intention can draw on limited capacity (Smith, 2003).
- A 2026 meta-analysis of 93 comparisons found a medium-sized effect of delay interval on time-based prospective memory, with better performance at shorter intervals, and a negative link with time-perception error across 18 comparisons (Manavi et al., 2026).
AI Measurement
Level of Measurement: Directly Measured
AI Memory Performance: Not Yet Compared
- PM-Bench, a peer-reviewed 2026 benchmark, adapts the psychologists' Virtual Week paradigm for AI agents (language models that take actions with tools): over a simulated seven-day week with 81 scored event- and time-based tasks, the agent must keep up an ongoing activity and act on earlier intentions when the right cue or time arrives (Liu & Gabriel, 2026).
- Across eight models, per-model scores (Set-F1, which penalizes both missed and unneeded actions) ranged from 42.0% (Qwen3-8B with no added support) to 79.1% (GPT-5.4 given optional periodic check-in prompts); averaged over models the best setup reached only 65.1%, and tasks needing clock-checking had a 54.7% hit rate (Liu & Gabriel, 2026).
- A 2026 preprint, TriggerBench, tests prospective memory by name against matched retrospective-memory controls: with the relevant constraint still inside the context window (what the model can see in the current conversation), accuracy at recalling it on request stayed at or above 98% up to 100,000 tokens, while spontaneously acting on it when the trigger appeared fell from above 90% to below 40% as the conversation lengthened (Zhang et al., 2026).
- Across nine model configurations in TriggerBench, prospective-memory accuracy spread from 40.7% (Qwen3-32B) to 83.2% (GPT-5.2 at high reasoning effort), but that best model scored only 43.9% on control trials where no reminder was due, a pattern resembling commission errors in people (Zhang et al., 2026).
- Saved memory did not help: with GPT-4.1, keeping the whole conversation in view scored 60%, versus 51.6% for retrieval from a stored archive (RAG), 53.2% for Mem0 and 41.4% for A-MEM, with only Letta-Sim, which pins key facts into the model's standing instructions, slightly higher at 62.3% (Zhang et al., 2026).
- A peer-reviewed 2026 workshop study treated formatting instructions as intentions to carry out while solving other tasks: across o4-mini, DeepSeek-V3.1 and Llama-3.3-70B and over 8,000 prompts, compliance fell 2-21% under concurrent task load (up to 50% for end-of-response actions), and a trailing reminder restored it to 90-100% in many settings, echoing ongoing-task load and cue-salience effects in people (Mittal, 2026).
- Not yet compared, because none of these studies tested people, though a human group could be given the same tasks: human studies report the proportion of cues acted on over minutes to days, whereas the AI benchmarks use different tasks, delays measured in words of conversation, and scores such as Set-F1 or model-judged accuracy, so there is no common scale; the AI failures are notable because the intention often remained visible to the model (Liu & Gabriel, 2026).
Memory Processes · Degradation
Forgetting
Definition
- Forgetting is the loss of access to stored content in a way that tracks how likely that content is to be needed; how available a memory stays follows the frequency, recency and spacing of past encounters, and records of the environment (newspaper headlines, speech to children, email) show the same statistical patterns, so forgetting is an adaptive bet rather than simple wear (Anderson & Schooler, 1991).
- Time-dependent decay is distinguished from interference: a regulated decay process, working mostly during sleep, removes selected memories, whereas interference arises when competing material is being encoded or retrieved; interference effects such as the fan effect and proactive or retroactive interference belong to a separate mechanism (Hardt et al., 2013).
- On this view transience is as useful as persistence: shedding outdated details keeps memory-guided decisions flexible in a changing world and keeps them from overfitting to particular past events (Richards & Frankland, 2017).
Neural Substrate
- In rats, blocking the removal of GluA2-containing AMPA receptors from hippocampal synapses during the retention interval prevented the normal forgetting of established long-term object-location memories without affecting learning, pointing to an active, decay-like erasure process (Migues et al., 2016).
- Work beginning in fruit flies identifies intrinsic forgetting, in which dopamine released onto memory-holding cells activates a Rac1/Cofilin pathway that slowly degrades memory traces; the authors propose this may be the brain's default state, competing with consolidation (Davis & Zhong, 2017).
- In adult mice, increasing the birth of new neurons in the hippocampal dentate gyrus after learning was enough to induce forgetting, and reducing it in infant mice lessened their rapid forgetting (Akers et al., 2014).
- In human fMRI, faces studied with spaced rather than massed repetitions were better recognized later, and spacing reduced repetition suppression in the fusiform gyrus, a neural signature linked to that advantage (Xue et al., 2011).
Human Measurement
Level of Measurement: Directly Measured
- The standard paradigm is the retention curve, dating to Ebbinghaus's savings method; a 2015 replication (one participant, 70 hours of learning and relearning lists after 20 minutes to 31 days) reproduced his curve of steep early loss followed by slower decline, with an upward jump at 24 hours (Murre & Dros, 2015).
- Pooling 210 published retention data sets and fitting 105 candidate functions, the best descriptions were logarithmic, power and related functions of time, in which loss is fast at first and then slows (Rubin & Wenzel, 1996).
- A meta-analysis of the spacing effect covered 839 assessments from 317 experiments in 184 articles and found that the gap between study sessions and the retention interval work together: the gap that gives the best retention grows as the retention interval grows (Cepeda et al., 2006).
- In more than 1,350 people taught facts, reviewed after a gap of up to 3.5 months and tested up to a year later, the best gap was about 20 to 40% of a 1-week test delay but only about 5 to 10% of a 1-year delay (Cepeda et al., 2008).
- The environmental side of the account has been measured directly: in 40 people keeping 100-day diaries of social contact, the chance of future contact with someone rose linearly with frequency of past contact and fell as a power function of time since last contact, the same shapes seen in memory (Pachur et al., 2014).
AI Measurement
Level of Measurement: Indirectly Measured
AI Memory Performance: Comparison Not Possible
- In training, a model's recall of a fact tracks how often it appeared: across GPT-Neo, BLOOM and GPT-3 models, question-answering accuracy correlated strongly with the number of relevant documents in the training data, and larger models helped only slowly with rarely seen facts; this is frequency, one half of the need signal, measured under another name (Kandpal et al., 2023).
- Using monthly Wikipedia page views as a stand-in for how often an entity comes up, ten language models were much less accurate on facts about less popular entities, and scaling up mainly improved recall of popular ones (Mallen et al., 2023).
- Tracking facts inserted during pretraining of OLMo models (1B and 7B) showed retention falling linearly with the logarithm of later training steps, the same curve family that fits human retention, and facts repeated verbatim were forgotten faster than paraphrased ones; no human data were placed on the same scale (Chang et al., 2024).
- Image, speech and language models were found to forget the specifics of training examples seen early in training, so examples seen recently are the ones most at risk of being recalled; this is a recency effect in training, but loss here is driven by further training rather than by how likely the content is to be needed (Jagielski et al., 2023).
- In the conversation (the context window, meaning the text the model can see right now), models used information best at the start or end of a long input: with 20 documents GPT-3.5-Turbo scored 53.8% when the answer sat in the middle, below its 56.1% with no documents at all; the authors link this U-shape to the human serial-position effect but tested no people (Liu et al., 2024).
- A 2023 preprint probing GPT-J found only a slow decline in recall as neutral filler text was added between a fact and the question, a clear benefit when a repetition was spaced rather than massed, and much larger losses from interfering facts; comparisons with human data were qualitative (Janik, 2023).
- In saved memory, agent systems build in need-like forgetting: MemoryBank weakens or reinforces stored memories using an Ebbinghaus-style curve based on elapsed time and significance, and Generative Agents rank memories partly by a recency score that decays by a factor of 0.995 per game hour since last use; neither paper tests whether this forgetting tracks actual need or human retention, so this place is unmeasured (Zhong et al., 2024; Park et al., 2023).
- The one same-test comparison of retention in the conversation gave 100 Prolific participants and GPT-2 Small (2019) text spans repeated up to three times; people saw only the previous 10 words when predicting, while the model saw everything, and it reached near-perfect prediction after one presentation while people improved modestly (human-model correlation fell from r = .87 to .05); the gap favours the side the test was easier for, so it is not counted (Vaidya et al., 2023).
- In an artificial-language study of spacing, three LLMs (including DeepSeek-V3 and Llama 3.1) outperformed people but, unlike people, did better with examples placed side by side than spaced apart, and effects varied by model; because the examples stayed in the prompt, the models' higher scores and missing spacing benefit reflect a test with nothing to forget, not a comparison of forgetting (Xiong & Wu, 2025).
- A 2025 preprint fitted the ACT-R decay parameter to in-context retention curves: four models ranged from d = 0.14 (Llama 3) to 0.35, all forgetting more slowly than a human reference of d = 0.50 (SD 0.05) taken from past modelling work, with Mamba and RWKV-7 closest in overall pattern; earlier material stayed in the context window and no human sample was tested, so this does not count either (Kang et al., 2025).
- The first-pass rating rested on a 2026 CogSci study of the fan effect in 112 people and retrieval systems; that measures interference between competing memories, a separate mechanism, so it is not used here (Guo et al., 2026).
- Comparison Not Possible: every human-AI test of this ability either let the AI keep looking at the material and found it did as well or better than people, or had no documented human sample, and no study places a model's training-based retention (where it cannot look back) against human retention curves or the environmental need statistics people track (Vaidya et al., 2023; Xiong & Wu, 2025; Kang et al., 2025).
Memory Processes · Degradation
Retrieval Interference
Definition
- The fan effect: the more separate facts people have learned about one concept (for example, several locations for the same person), the slower and less accurately they retrieve any single one of those facts, because the facts compete for the same retrieval cue (Anderson, 1974).
- Interference theory separates proactive interference, where earlier learning impairs memory for later material, from retroactive interference, where later learning impairs memory for earlier material, and treats competition from other stored material, rather than passive decay, as a main cause of forgetting in laboratory list learning (Underwood, 1957).
- Cue-overload interference, as in A-B, A-C list learning where one cue is tied to competing responses, is distinct from consolidation-based interference, in which any new mental activity disrupts a recent memory that has not yet stabilised, even when the new material is unrelated (Wixted, 2004).
Neural Substrate
- Across imaging studies, most using the recent-probes task (rejecting a probe that belonged to the previous trial's memory set), resolving proactive interference repeatedly engages left inferior frontal cortex (Jonides & Nee, 2006).
- In fMRI, recent negative probes slowed responses, raised errors, and increased activation in left mid-ventrolateral prefrontal cortex and bilateral frontopolar cortex, suggesting more than one control process resolves proactive interference (Badre & Wagner, 2005).
- In an fMRI fan-effect study, people who had mastered person-location pairs with one to three associations each were slower to recognise pairs as associations increased, and prefrontal cortex tracked this competition while posterior parietal cortex was active during retrieval but did not (Sohn et al., 2003).
- A patient with a semantic short-term memory deficit showed proactive interference on the recent-negatives task well outside the range of healthy controls, which the authors attribute to abnormal persistence of earlier material rather than rapid decay (Hamilton & Martin, 2007).
Human Measurement
Level of Measurement: Directly Measured
- Proactive interference is measured with a standard paradigm and pooled: a meta-analysis of 26 studies (82 data sets) found it shrinks between ages 4 and 13, and in 125 children and college students recall of three-word sets was accurate on trial 1 but fell over trials 2 to 4 (Kail, 2002).
- Proactive and retroactive interference indices from list-learning tasks are standard enough to be meta-analysed: across 20 studies, adults with ADHD did not differ from controls on proactive interference (g = 0.13, 95% CI -0.02 to 0.28), while people with ADHD showed slightly more retroactive interference (g = 0.17) (Orban et al., 2022).
- The person-location fan effect replicated in 134 adults (46 in the lab, 88 online) with Dutch materials: sentences whose location had four associations were recognised less accurately than those with two (78% vs 83%) (Link et al., 2026).
- Fan effects on accuracy depend on how well material is learned: in Wickelgren and Corbett's data from six people who saw word pairs or triples once for 3 s, fitted maximum d' (a discrimination score where 0 is chance) was 4.01 for fan 1 and 2.72 for fan 2, whereas ten people who learned person-location facts over five sessions showed no difference in maximum accuracy, only a slower rise (Schneider & Anderson, 2012).
- In a 2026 study of 112 online adults who studied 60 person-location sentences for 5 s each, then did a 2-minute arithmetic task and an old/new recognition test, accuracy fell from 91.2% at fan 1 to 82.4% at fan 2 and 71.8% at fan 4 (Cohen's d = 1.12) (Guo et al., 2026).
AI Measurement
Level of Measurement: Directly Measured
AI Memory Performance: Comparable to Humans approximate provisional
- The fan paradigm has been given to AI by name in all three places AI holds information: in the context window (what a model can see in the current conversation) and in training data, where a 2024 preprint found fan effects in the confidence of only 2 of 6 language models (Mistral and SOLAR), and in saved memory, where a 2026 study tested retrieval systems against people (Roberts et al., 2024).
- Deciding study: a 2026 paper (accepted at CogSci 2026 per its arXiv listing) gave three retrieval systems, the programs that look up stored text for an AI model, the same person-location sentences people studied: DPR (2020), ColBERTv2 (2022) and the hippocampus-inspired HippoRAG (2024), scored correct when the studied sentence came back first, at fans of 1, 2, 4 and 8 (only its fan and competitor results bear on this mechanism; its serial-position results bear on a different one) (Guo et al., 2026).
- On a fitted interference slope (how fast discrimination falls as fan grows, on a log scale; lower is better), people scored 0.41 (95% CI 0.35 to 0.47), HippoRAG 0.44 (0.37 to 0.51), ColBERTv2 0.53 (0.45 to 0.61) and DPR 0.67 (0.58 to 0.76), so systems spanned 0.44 to 0.67 and only DPR differed reliably from people; in raw accuracy DPR fell from 94% to 58% by fan 4 while HippoRAG tracked the human curve (about 84% and 72% at fans 2 and 4) (Guo et al., 2026).
- This comparison is approximate and favoured the AI: the systems held every sentence verbatim and were queried with no delay, while people had to remember after a distractor, and people judged old versus new while systems ranked stored sentences; HippoRAG matching people despite that advantage is a tie, not a win (Guo et al., 2026).
- The Guo et al. numbers need independent checking: the reported slopes are about half what the paper's own human d' values (2.84, 2.21, 1.67) imply, and a query of a person's name alone, as described, could not yield the above-chance retrieval scores reported, so what the systems were asked is unclear; the study was not preregistered (Guo et al., 2026).
- A second, weaker comparison in the context window: a 2025 preprint gave GPT-4 (a 2024 version) a 32-pair person-location recognition test with the study list in the same prompt, and d' fell from about 4.1 at fan 1 to 3.1 at fan 2; smaller open models (Mistral-7B, Llama-3-8B) showed the same trend at lower accuracy (Cao et al., 2025).
- Cao et al.'s human values (about 3.7 and 2.5) are not stated in their cited source's text but match its plotted long-lag d' for Wickelgren and Corbett's six people studying word pairs and triples under response deadlines; with different materials, no reported spread and GPT-4 able to see the list, its similar cost (about 1.0 vs 1.2 to 1.3 d' units) is only a tie (Schneider & Anderson, 2012).
- Proactive interference tests in AI have no human group: in a 2025 preprint (ICML workshop), dozens of language models tracking 46 repeatedly updated keys lost accuracy log-linearly toward zero as overwritten values piled up, even with the answer just before the question; the paper states that people do well on such tasks but tested none (Wang & Sun, 2025).
- An A-B, A-C style study of 39 language models found proactive interference exceeded retroactive interference in every model (d = 1.73), which the authors call the reverse of the human pattern, but that contrast rests on past literature rather than people tested on the same items, so it cannot be scaled (Chattaraj & Raj, 2026).
- Where people and models did share an updating task, a 2026 preprint testing 50 online adults and 9 language models on ten memory tasks, including tracking changing person-city pairs, found models near ceiling on every task with the material still in view, which does not count and measures updating capacity rather than an interference cost (Wang et al., 2026).
Memory Processes · Degradation
Motivated Forgetting
Definition
- Motivated forgetting is forgetting brought about by a person's own goal to keep an unwanted memory out of awareness, achieved by recruiting the same executive control used to stop habitual actions; unlike passive decay or interference, the forgetting grows with the number of times the memory is actively avoided (Anderson & Green, 2001).
- Suppression can act at encoding (a cue to forget an item or list just studied, as in directed forgetting) or at retrieval (stopping recall when a reminder appears, as in think/no-think), and both are thought to leave lasting biases in which experiences stay accessible (Anderson & Hanslmayr, 2014).
- Directed forgetting comes in two procedures that are not interchangeable: the item method, where a forget cue follows each item and the effect mainly reflects reduced rehearsal at encoding, and the list method, where the cue follows a whole list; item-method effects are reliably larger (Titz & Verhaeghen, 2010).
- Within retrieval suppression, direct suppression (blocking the memory) is distinct from thought substitution (deliberately recalling a replacement memory); the two produce forgetting through opposite routes (Benoit & Anderson, 2012).
Neural Substrate
- In fMRI, direct suppression reduced hippocampal activity via inhibition originating in right dorsolateral prefrontal cortex, whereas thought substitution engaged left caudal and mid-ventrolateral prefrontal cortex to retrieve substitute memories (Benoit & Anderson, 2012).
- Combining fMRI with magnetic resonance spectroscopy, higher resting GABA concentration in the hippocampus (not prefrontal cortex) predicted better suppression-induced forgetting and stronger fronto-hippocampal coupling, implicating local inhibitory interneurons (Schmitz et al., 2017).
- Repeated suppression of aversive scenes reduced their later neural reactivation in parahippocampal cortex, and reduced reinstatement tracked a loss of vividness, suggesting suppression degrades the stored representation (Meyer & Benoit, 2022).
- A neuroimaging meta-analysis across intentional-forgetting paradigms found a mainly right-lateralized frontoparietal core network active regardless of paradigm, plus a supporting frontal-hippocampal network (Gamboa et al., 2022).
Human Measurement
Level of Measurement: Directly Measured
- The think/no-think paradigm, built to test retrieval suppression, has been pooled in a 2026 meta-analysis of 500 effects from 120 studies, finding a consistent suppression-induced forgetting effect of about d = 0.20 to 0.40, larger with more suppression repetitions and for autobiographical memories (Clark et al., 2026).
- An earlier focused meta-analysis (96 effects, 25 studies) estimated suppression-induced forgetting in healthy adults at 0.28 (95% CI 0.14 to 0.43), with the effect absent or reversed in more anxious or depressed groups (Stramaccia et al., 2021).
- The effect is not always reproduced: a preregistered online replication with 150 healthy adults found no suppression-induced forgetting (same-probe d = 0.03, 95% CI -0.13 to 0.20) (Wiechert et al., 2023).
- Directed forgetting is larger and more consistent: a meta-analysis found forget cues reduced memory by d = 1.17 in younger adults and d = 0.81 in older adults, with bigger effects for the item method and for recall than recognition (Titz & Verhaeghen, 2010).
- A meta-analysis of 36 item-method directed forgetting studies found a reduced directed forgetting effect in clinical groups, particularly anxiety and depression, relative to controls (Pevie et al., 2023).
AI Measurement
Level of Measurement: Indirectly Measured
AI Memory Performance: Comparison Not Possible
- No study located gives a language model the directed forgetting or think/no-think procedure, or tests people and AI on the same forget-on-instruction task; AI work instead appears under other names (in-context unlearning, machine unlearning, memory deletion), so the AI level is indirect (Wang & Sun, 2025).
- Current conversation: in a 2025 workshop paper, 30+ language models tracked repeatedly updated values in their context window (what the model can see in the conversation); explicitly telling models to forget earlier values improved accuracy by under 10 percentage points, and all models declined as outdated values piled up, with larger models declining more gradually (Wang & Sun, 2025).
- Current conversation: models fine-tuned to forget named facts when asked in the prompt (Llama 2 and Mistral) reached up to 95% forget accuracy while retaining 80% of unrelated knowledge, but the correct answer was still present in middle layers, so the models 'pretend to forget' rather than lose the information (Takashiro et al., 2025).
- Training: 'unlearning' methods meant to remove knowledge from a model's weights (its learned parameters) left it largely recoverable, with brief retraining on related facts restoring 88% of pre-unlearning accuracy (2024 preprint), a pattern reminiscent of suppressed but still-stored human memories (Deeb & Roger, 2024).
- Saved memory: a 2026 preprint on one model family (Gemma 3) found that deleting an exchange from the assistant's stored memory disclosed fewer deleted answers than simply instructing the model to forget, yet an attack could still tell the edited memory from one that never held the record (Ramesh, 2026).
- Agent memory benchmarks test forgetting outdated facts when they are replaced (selective forgetting in MemoryAgentBench, a 2025 preprint), finding that no tested system masters it, but they include no human baseline (Hu et al., 2025).
- Comparison is not possible because there is no common scale: human studies measure the drop in recall for items people were cued to forget or suppress, while AI studies measure intrusions from outdated information or leakage of deleted facts, and none reports human data on its task (Clark et al., 2026).
Metacognition · Attribution
Source Monitoring
Definition
- Source monitoring is the set of processes by which people judge where a memory came from, such as who said something, where it was read, or whether it was seen or only thought about; the source is not stored as a label but inferred at retrieval from the memory's perceptual, contextual and cognitive-operation details, using flexible decision criteria that can err (Johnson et al., 1993).
- Reality monitoring is the subtype that separates internally generated experiences (imagined, thought or inferred) from externally perceived ones, whereas external source monitoring discriminates between two outside sources, such as two speakers (Johnson & Raye, 1981).
- Source monitoring is distinct from remembering the content itself: in the eyewitness misinformation paradigm, people who could identify where a misleading suggestion came from still sometimes attributed it to the original event, and the suggestibility effect disappeared when the test directed attention to sources (Lindsay & Johnson, 1989).
Neural Substrate
- Patients with frontal lobe lesions recalled newly learned facts as well as matched controls but often attributed those facts to the wrong source, implicating prefrontal cortex in binding facts to the context in which they were learned (Janowsky et al., 1989).
- In event-related fMRI, encoding activity in the hippocampus and posterior parahippocampal cortex predicted later source recollection, whereas perirhinal activity tracked later item recognition, separating medial temporal routes to source and item memory (Davachi et al., 2003).
- A meta-analysis of 51 fMRI studies found convergent source-memory activity centred on left lateral prefrontal cortex (middle and inferior frontal gyri) for external source monitoring, and on left inferior parietal lobule and angular gyrus for reality monitoring, suggesting partly distinct networks (Martín-Luengo et al., 2026).
- Among 53 healthy adults, those lacking the paracingulate sulcus (a fold in medial anterior prefrontal cortex) in both hemispheres showed significantly poorer reality monitoring and poorer metacognitive insight into their own accuracy (Buda et al., 2011).
Human Measurement
Level of Measurement: Directly Measured
- The standard source-monitoring test presents items from two or more sources (two speakers, seen versus imagined) and later asks for an old/new judgment plus a source judgment; multinomial processing tree models separate detecting that an item is old from discriminating its source (Batchelder & Riefer, 1990).
- In a representative two-speaker test with 117 adults, participants attributed about 71% to 73% of objects to the correct speaker, where chance was 50% (Beck & Weber, 2024).
- A meta-analysis of 90 studies (3,197 younger and 3,192 older adults) found that age impairs memory for source and other associative information more than memory for the items themselves, especially under intentional learning (Old & Naveh-Benjamin, 2008).
- A meta-analysis of 53 studies (85 experiments, N = 3,040) found that emotional content slightly lowers source accuracy relative to neutral content (d = -0.14 for negative, -0.11 for positive), while higher arousal improves it (d = 0.27 to 0.49) (Pereira et al., 2023).
- Across 67 behavioural, signal-detection and imaging trials (2,297 people with psychosis, 1,745 healthy controls), source-monitoring accuracy was substantially lower in psychosis (d = -0.83), showing the task reliably separates groups (Damiani et al., 2024).
AI Measurement
Level of Measurement: Directly Measured
AI Memory Performance: Not Yet Compared
- A 2026 preprint gave six open-weight language models the reality-monitoring word-pair paradigm from human research (Johnson & Raye), in which the second word is either shown or generated, followed by an 'internal or external?' source judgment; all models identified their own generated words with 100% accuracy when that word was still in the context window, meaning what the model can see in the current conversation (Ranjan et al., 2026).
- Systems differed widely in the same preprint: on shown words, accuracy ranged from 47.9% (Gemma 3 12B) to 94.4% to 100% (Llama 3.3 70B); with a delayed source test after 20 or 40 trials, most models discriminated sources at chance (d' about 0), while the best, Llama 3.3 70B, reached d' = 2.01 and 1.62 (Ranjan et al., 2026).
- No comparison can be drawn from this study: its human reference is an anonymised prior dataset described only qualitatively (no sample size or scores), and the earlier trials stayed in the model's conversation history at test, so the best model's strong result favours the side that could still look back at the material (Ranjan et al., 2026).
- In a benchmark of 50 very long two-person conversations, long-context models were especially prone to assigning events to the wrong speaker, scoring as low as 2.1 F1 on trick questions versus 70.2 for GPT-4-turbo; a human score of 89.4 is reported without the number of people or procedure, so it cannot serve as a documented baseline (Maharana et al., 2024).
- For saved memory, a 2026 preprint built a source-monitoring probe for a robot's stored memory (for example, whether it learned about an object from a scene description or an object detector); its structured memory scored 78.6 of 100 on an AI-judged scale, and simpler memory designs fell 20 to 35 points behind on matched probes, with no human data (Rasheed & Kabtoul, 2026).
- For training, a 1.1-billion-parameter model trained on synthetic biographies with document labels answered 88.8% of questions on unseen documents correctly but named the correct source document only 47.3% of the time, a gap between content and source memory; the ability required special training, since standard models are not trained to track sources (Khalifa et al., 2024).
- For retrieved material, a human audit of four AI search engines found that only 51.5% of generated sentences were fully supported by their citations and 74.5% of citations supported the sentence they were attached to, so many claims were not backed by the source they were credited to (Liu et al., 2023).
Metacognition · Attribution
Reality Monitoring
Definition
- Reality monitoring is the set of judgment processes people use to decide whether a memory came from the outside world (something perceived) or from their own minds (something thought, imagined or planned). It is an attribution made at remembering, not a separate memory store (Johnson & Raye, 1981).
- In the wider source monitoring framework, reality monitoring (internal versus external origin) is distinguished from external source monitoring (which of two outside sources, such as two speakers) and internal source monitoring (which of two of one's own acts, such as said versus thought) (Johnson et al., 1993).
- The judgment relies on the qualities of the memory itself: memories of perceived events carry more perceptual and contextual detail, and supporting memories, whereas people justify memories of imagined events more by reasoning from prior knowledge (Johnson et al., 1988).
- Reality monitoring errors range from everyday confusions between real and imagined experiences to clinical symptoms such as hallucinations, which separates it from simple recognition (whether an item is old or new) and from judgments about how well something is known (Simons et al., 2017).
Neural Substrate
- In 53 healthy adults selected by MRI, those lacking the paracingulate sulcus (a fold in the medial anterior prefrontal cortex) in both hemispheres showed reduced reality monitoring accuracy and poorer metacognitive insight into their own performance (Buda et al., 2011).
- In 153 people with schizophrenia, hallucinations were linked to a shorter paracingulate sulcus, with each 1 cm reduction raising the likelihood of hallucinations by 19.9%, tying reality monitoring failure to the same medial prefrontal structure (Garrison et al., 2015).
- A coordinate-based meta-analysis of fMRI reality monitoring studies (9 studies, 172 healthy adults) found activity in right anterior medial prefrontal cortex, cerebellar lobule VI and anterior thalamic projections, though few clusters survived strict correction (Lavallé et al., 2023).
- A 2026 fMRI meta-analysis (17 reality monitoring and 34 external source monitoring studies) found shared activity in left middle frontal gyrus but a reality-monitoring cluster centred on the left inferior parietal lobule and angular gyrus (Martín-Luengo et al., 2026).
Human Measurement
Level of Measurement: Directly Measured
- The standard test has people see a complete word pair (e.g., "Laurel and Hardy") or imagine the missing second word, then later shown the first word and asked whether its partner was perceived, imagined, or whether the word is new (Garrison et al., 2017a).
- In 47 healthy adults tested on familiar word pairs in six short study-test blocks, source accuracy among items correctly recognised as old was 85% (SD 7%) in low and 84% (SD 8%) in high hallucination-prone groups, with no group difference; the paper reports no signal-detection (d′) score for this task (Garrison et al., 2017b).
- Healthy adults' errors lean outward: in 45 participants, imagined items called perceived outnumbered perceived items called imagined (partial eta squared = .40 among the 33 who made errors), and imagining in the auditory rather than visual modality lowered accuracy for imagined items (d = 1.11) (Garrison et al., 2017a).
- When the imagined and perceived words come from the same false-memory (DRM) word lists, performance collapses: 34 adults judging 72 perceived, 72 imagined and 72 new items identified imagined sources correctly only 29% of the time (SD 26%, chance 33%), calling most of them perceived (Ranjan & Odegaard, 2024).
- A meta-analysis of 44 studies (1,566 patients, 1,175 controls) found source monitoring reduced in psychosis (SMD = 0.46), with the largest deficit for imagined stimuli (accuracy SMD = 0.98) (Damiani et al., 2022).
- An earlier meta-analysis of 24 clinical and non-clinical studies found moderate-to-large externalizing biases in people who hallucinate or are prone to hallucinations, robust in source monitoring and signal detection tasks (Brookwell et al., 2013).
AI Measurement
Level of Measurement: Directly Measured
AI Memory Performance: Not Yet Compared
- A 2026 preprint gave six open-weight large language models released in 2024–2025 (Gemma 3 12B and 27B with and without quantization-aware training, Llama 3.3 70B, Llama 4 16×17B) the perceived-versus-imagined word-pair task by name, using word pairs drawn from the same DRM lists as a human study. When the model's own generated word was still in view in the context window (what the model can see in the current conversation), all six scored 100% on self-generated items, which the authors attribute to simply rereading it (Ranjan et al., 2026).
- In the delayed version (all pairs encoded first, source tested afterward), models without feedback averaged 76% correct on perceived items but 38% on imagined ones. The spread across systems was wide: most models showed no real discrimination (d′ ≈ 0), while Llama 3.3 70B reached d′ = 2.01 with 20 trials and 1.62 with 40, roughly 70–90% correct on both sources in the paper's figure (Ranjan et al., 2026).
- People on the same word lists did far worse than that best model: in the human study (N = 34, incidental encoding, one session) accuracy was 63% (SD 18%) for perceived, 29% (SD 26%) for imagined and 66% (SD 21%) for new items, and a model of their responses put source discrimination at only 0.11 on a 0–1 scale. These figures come from the authors' 2023 PsyArXiv revision preprint; the journal full text could not be retrieved (Ranjan & Odegaard, 2024).
- The tests were not the same and every difference favoured the models: people chose among perceived, imagined and new for 72 items per source from memory, whereas models chose between only two options, faced no new items, had 5 or 10 items per source, and kept the whole encoding transcript, including which second words had been supplied or left blank, in their context window at test (Ranjan et al., 2026).
- Not yet compared: people and models have both done this task on the same word lists, and a matched version could be run, but no study reports human and model results on a common scale (humans as three-choice accuracy and model parameters, models as two-choice accuracy and d′), and the authors themselves call their human comparison qualitative. The only visible direction, the best model outscoring people, is the direction the easier test favoured, so it does not count (Ranjan et al., 2026).
- The familiar-word-pair human results cannot stand in: their 85% is source accuracy among items recognised as old on a three-choice test with different materials, no d′ was published for it, and on the models' own word lists people scored far lower, so any placement of the models within that range would rest on a conversion no paper has made (Garrison et al., 2017b).
- A 2025 Anthropic report (not peer reviewed) found Claude models could sometimes tell internally planted "thoughts" (injected activation patterns) from text they were reading, with the best, Opus 4.1, detecting injections in about 20% of trials and no false alarms in 100 control trials, and could disavow words prefilled in their name unless the matching internal state was also planted; older Claude models did less well, and no human baseline was used (Lindsey, 2025).
- Related self-attribution tests also lack human baselines: GPT-4 picked out its own summaries 73.5% of the time while a study of ten leading models found no consistent self-recognition, and a 2026 preprint on retrieval systems (models that look up saved documents) found internal signals, not the models' own judgments, revealing whether an answer came from training or the retrieved text; no study tests reality monitoring over assistants' saved cross-session memory (Panickssery et al., 2024; Davidson et al., 2024; Yu et al., 2026).
Metacognition · Confidence
Feeling of Knowing
Definition
- The feeling of knowing (FOK) is the judgment, made after failing to recall an answer, that the answer is nonetheless stored and would be picked out on a later recognition test; it was introduced with the recall–judgment–recognition procedure, in which people try to recall, rate their FOK, and then take a recognition test (Hart, 1965).
- The accessibility account holds that FOK is not direct access to the stored answer but an inference from how much partial information, correct or incorrect, comes to mind during the retrieval attempt, which explains both why FOK is usually accurate and when it misleads (Koriat, 1993).
- A separate, rapid 'initial' FOK is made before any retrieval attempt, when people decide within moments whether an answer can be retrieved or must be worked out; it is driven by familiarity with the question's terms rather than with the answer (Reder & Ritter, 1992).
- FOK is a prediction of upcoming memory performance, which separates it from retrospective confidence about an answer already given; the two dissociate across memory tasks and in memory-impaired groups (Meunier-Duperray et al., 2025).
Neural Substrate
- In 14 patients with frontal lesions, FOK accuracy was markedly impaired while confidence in their own recall stayed accurate, and lesion overlap pointed to right medial prefrontal cortex (Schnyer et al., 2004).
- Five patients with focal medial prefrontal damage, compared with 19 controls, made less accurate FOK judgments even when their memory performance improved on an easier task, whereas their judgments of learning were intact (Modirrousta & Fellows, 2008).
- An fMRI study of the recall–judgment–recognition task found left dorsolateral, left anterior, bilateral inferior and medial prefrontal activity that rose with FOK strength, and right inferior frontal regions engaged by FOK but not by successful recall (Kikyo et al., 2002).
- In event-related fMRI of episodic memory, trials with a feeling of knowing produced frontal and left parietal activity intermediate between successful recall and failed recall without a feeling of knowing (Maril et al., 2003).
Human Measurement
Level of Measurement: Directly Measured
- The standard test is the recall–judgment–recognition procedure with general-knowledge questions (semantic FOK) or studied word pairs (episodic FOK); every study in the one meta-analysis scored accuracy as the within-person gamma correlation between FOK ratings and later recognition (Devaluez et al., 2023).
- That meta-analysis of 20 studies (922 younger, 966 older adults) found no age difference in semantic FOK accuracy (g = -0.10 [-0.29, 0.10]) but lower episodic FOK accuracy in older adults (g = 0.53 [0.28, 0.78]), partly explained by their weaker recognition memory (Devaluez et al., 2023).
- In 50 young adults recalling names for faces, with 0–100 FOK ratings and four-choice recognition, gamma for unrecalled items was 0.56 (SD 0.40) for famous faces (semantic) and 0.39 (SD 0.32) for newly learned face–name pairs (episodic) (Eakin et al., 2014).
- Gamma equals twice the area under the ROC curve minus one (the area being the probability that a correct item gets a higher rating than an incorrect one); the usual formula for gamma drops tied ratings and so runs further from zero than this ROC-based value, and the authors recommend the ROC version (Higham & Higham, 2019).
- In 38 young adults rating FOK for all items of a face–scene task before three-choice recognition, usual gamma was 0.28 to 0.38 (SD 0.42 to 0.51) across three trials, but ROC-based gamma was only 0.14 to 0.26 (SD 0.30 to 0.38), equivalent to an area under the ROC curve of about 0.57 to 0.63 (Yeung, 2022).
- Accuracy depends heavily on procedure: in 240 mostly online adults, FOK efficiency (meta-d'/d') was 0.06 to 0.07, against 0.55 to 0.65 in the same group's lab study of 92 students, apparently because participants rated whether they had recalled rather than predicting recognition (Meunier-Duperray et al., 2025).
AI Measurement
Level of Measurement: Indirectly Measured
AI Memory Performance: Not Yet Compared
- Not yet compared: no study found through September 2026 gives an AI the recall–judgment–recognition procedure (fail to produce an answer, rate whether it would be recognized, then take a recognition test), with or without a human baseline, so there are no AI results on the ability as psychologists measure it (Hart, 1965).
- A 2026 preprint that uses the FOK name told seven language models not to solve each question but to rate how likely they were to answer it correctly; on 720 factual test questions, discrimination (area under the ROC curve, where 0.5 is chance) ranged from 0.56 (Qwen3-7B) to 0.67 (Qwen3-32B), with Claude Sonnet 4.6 at 0.66 and GPT-5.4 at 0.57, and no human baseline (Kale, 2026).
- That task is a different measure: it predicts the model's own recall from the question alone, with no failed-recall step and no recognition test, so it resembles the rapid initial FOK or a prospective confidence rating, and its score mainly reflects telling answerable from unanswerable questions, which is exactly what FOK for unrecalled items leaves out; people who rate recall instead of predicting recognition score very differently, so the two cannot stand in for each other (Meunier-Duperray et al., 2025).
- The scale conversion itself is published (gamma = 2 × area − 1), and one human FOK study reports the ROC-based value: 0.14 to 0.26 against the best model's 0.34, inside the human spread; this is not used, because the tests measure different judgments and the direction of any advantage is unclear (the models were told not to attempt retrieval, which supplies the partial information human FOK relies on, but were also not restricted to unrecalled items) (Yeung, 2022).
- The earliest close test trained a 52-billion-parameter model to estimate, from the question alone, the probability that it knows the answer; this separated known from unknown trivia questions well (0.864), transferred unevenly to other tasks (0.606 to 0.928), and rose when relevant text was placed in its context window, meaning what it can see in the current conversation (Kadavath et al., 2022).
- The gap FOK depends on does exist in models: in three 7–9-billion-parameter models, an internal readout ranked the correct answer above every wrong one for 9% of questions whose answer the model never produced in 1,000 attempts, and internal knowledge exceeded expressed knowledge by 14% to 57% across models; the models were not asked to sense this themselves (Gekhman et al., 2025).
- For saved or retrieved memory, systems use a model's sense of what it already knows to decide when to look information up in an external store, but they are scored on final question-answering accuracy, not on how well that sense predicts later recognition (Wang et al., 2023).
- Same-question human–AI studies exist only for confidence given after answering, which bears on retrospective confidence rather than FOK: on trivia, 110 people's ratings discriminated right from wrong answers at an area of 0.50 and four chatbots (each run 100 times) at 0.48 to 0.65 (Cash et al., 2025).
- Likewise, on Japanese general-knowledge questions rated after each answer, 87 people discriminated right from wrong answers at an area of 0.76 (SD 0.14) and GPT-4, run 87 times, at 0.63 (SD 0.20); again a retrospective confidence result, not FOK (Yoshizawa et al., 2026).
Metacognition · Confidence
Confidence Calibration
Definition
- Confidence accuracy has separable parts: metacognitive bias (calibration), meaning whether confidence runs higher or lower than accuracy, as in over- or underconfidence, and metacognitive sensitivity (resolution), meaning how well confidence separates one's correct from incorrect answers; efficiency is sensitivity relative to how well the task itself was done (Fleming & Lau, 2014).
- In the self-consistency account, confidence in an answer reflects how consistently that answer is supported across the representations retrieved from memory, so confidence tracks reliability rather than correctness, and overconfidence arises when a wrong answer is consistently favoured (Koriat, 2012).
- Retrospective confidence in answers already given is distinct from prospective predictions of success, such as feeling-of-knowing or judgments of learning: the two draw on different cues even when elicited in the same task, although overall overconfidence is shared across them (Fleming et al., 2016).
Neural Substrate
- A review of patient studies found that frontal-lobe integrity is closely linked to metamemory accuracy, that combined frontal dysfunction and poor memory severely impairs it, and that metamemory judgments can dissociate from memory retrieval itself (Pannu & Kaszniak, 2005).
- A coordinate-based meta-analysis of 47 neuroimaging studies found a domain-general network for confidence in one's own performance, spanning medial and lateral prefrontal cortex, precuneus and insula, for both memory and decision tasks (Vaccaro & Fleming, 2018).
- In macaques rating confidence in their recognition memory, reversibly inactivating prefrontal area 9/46d impaired confidence judgments for remote memories and area 6 for recent ones, while recognition itself stayed intact (Miyamoto et al., 2017).
- Patients with anterior prefrontal lesions (n = 7) lost metacognitive accuracy for perceptual decisions but kept it for memory decisions, compared with temporal-lobe patients (n = 11) and healthy controls (n = 19), pointing to partly domain-specific substrates (Fleming et al., 2014).
Human Measurement
Level of Measurement: Directly Measured
- The standard paradigm asks general-knowledge questions, often with two alternatives, followed by a confidence rating; a quantitative review of such studies found that the classic overconfidence and hard-easy effects largely disappear when items are sampled representatively and scale-end and regression artifacts are controlled (Juslin et al., 2000).
- A meta-analysis of 213 datasets in the Confidence Database found an across-person correlation of R = .22 between average confidence and average accuracy, much stronger for memory tasks than for perceptual tasks (Jin et al., 2022).
- For judgments of learning, a prospective relative of confidence, a meta-analysis of 112 effect sizes from 4,554 participants found that delaying the judgment greatly improves resolution (g = 0.93), consistent with confidence drawing on diagnostic cues from long-term memory (Rhodes & Tauber, 2011).
- In an eyewitness version with 1,200 adults, confidence was calibrated with accuracy for people who picked someone from the lineup but not for those who rejected it (Brewer & Wells, 2006).
- A synthesis of three decades of eyewitness research concluded that under pristine testing conditions confidence and accuracy are strongly related and high-confidence identifications are highly accurate, whereas unfair lineups and similar conditions break the relationship (Wixted & Wells, 2017).
AI Measurement
Level of Measurement: Directly Measured
AI Memory Performance: Comparable to Humans provisional
- A peer-reviewed 2026 study gave 87 Japanese-speaking adults (recruited on X, answering an online form in September–October 2023) and 87 separate runs of GPT-4 the same 60 general-knowledge questions about Japan as open-ended recall, four-choice and two-choice items, each followed by a 0–100% confidence rating; GPT-4 was queried in December 2025 through OpenAI's programming interface under the setting "gpt-4", which points to a June 2023 version, one question per fresh request, so it answered only from what it learned in training (Yoshizawa et al., 2026).
- GPT-4 answered more questions correctly than people (78–88% vs 52–73% across formats) but reported 93–97% confidence throughout, overconfident by 8–15 percentage points, while people's average confidence was within 4 points of their accuracy; individual people varied widely (SD of over- or underconfidence 15–16 points, computed from the authors' public data), so GPT-4's bias sits within the human range, about 0.6–0.75 SD above the human mean (Yoshizawa et al., 2026).
- For resolution, scored as type-2 AUROC (the chance that a correct answer drew higher confidence than a wrong one, .5 = chance), GPT-4 matched people on two-choice (.73 vs .70, SD .14) and open-ended recall (.78 vs .81, SD .12) but fell below chance on four-choice items (.40 vs .76, SD .13), where it was more confident in its wrong answers (99%) than its right ones (96%); across questions, people's confidence tracked item difficulty far more closely than GPT-4's (r = .83 vs .33) (Yoshizawa et al., 2026).
- The study's efficiency measures are information-based (meta-I and meta-I2r), not meta-d′ or M-ratio, and they do not line up with its AUROCs: on open-ended recall AUROC shows no reliable difference while both information measures favour people, and a mutual-information score counts GPT-4's inverted four-choice confidence as positive information; the text also reverses the direction of a format effect shown in the authors' own statistical output, and its "All" AUROC averages the three format scores rather than pooling trials (Yoshizawa et al., 2026).
- A peer-reviewed five-study paper (online 2025, in print 2026) gave online US adults and public chatbots (ChatGPT, Bard/Gemini and, in Studies 4–5, Claude Sonnet and Haiku, 2023–2025 versions, each run repeatedly through its chat interface) identical items and confidence scales, but its items were NFL and Oscar forecasts, sketch guessing, 22 more-than/fewer-than trivia questions built to require piecing facts together (110 adults), and 33 questions about an unnamed university's unpublished records (110 adults), so no study isolates confidence in answers recalled from memory (Cash et al., 2026).
- On the trivia questions the best system, Claude 3.5 Sonnet (released 2024), missed its own prospective score by 2.15 items versus 3.51 for people (SD 3.15), with AUROC .66 versus .55 (SD .10), while systems ranged from 2.15 to 4.94 items and .57 to .66; on university questions Claude 3.7 Sonnet (2025) missed by 1.08 items versus 4.84 (SD 3.74), yet people discriminated right from wrong answers better (.57 vs .53), and over all 24 calibration comparisons LLMs were better 13 times, equal 4 and worse 7 (Cash et al., 2026).
- Both are direct comparisons and both put AI at or above people, so the rating is Comparable: in the forecasting, trivia and records studies the best systems matched or beat people's calibration, and on general-knowledge recall GPT-4 was more accurate than people and within the human range on calibration and on two of three resolution formats. The important caveat is overconfidence: GPT-4 was 8–15 points more confident than it was accurate where people were within 4, and on four-choice items it was more confident in its wrong answers than its right ones. The rating is provisional because the recall study tests one 2023 system whose 87 runs were near-duplicates, and the other study's items are forecasts and estimates rather than answers from memory (Yoshizawa et al., 2026).
- Neither study examined saved memory; the closest test, a benchmark of chat assistants' long-term memory of past sessions, scores whether they decline questions their memory cannot answer (abstention), not whether confidence tracks accuracy, and has no human comparison (Wu et al., 2025).
Metacognition · Control
Metacognitive Control
Definition
- Metamemory is split into monitoring, where information flows up from memory to a person's model of it (judgments such as feelings of knowing), and control, where that model acts back on memory by starting, continuing or stopping activities such as study or retrieval (Nelson & Narens, 1990).
- Given free choice over study, people did not simply give most time to the hardest items, as discrepancy-reduction models predicted; they studied easy items first and spent most time on medium-difficulty items, which the region of proximal learning framework explains as study targeted at what is almost learned (Metcalfe & Kornell, 2003).
- Control acts on what people believe they know, not on what they actually know: when judgments of learning were made to diverge from later recall, restudy choices followed the mistaken judgments (Metcalfe & Finn, 2008).
- Control can be separated from monitoring: dividing attention while studying word pairs left the resolution of judgments of learning largely intact but impaired the link between those judgments and restudy choices, and the benefit of those choices (Peng & Tullis, 2021).
Neural Substrate
- In patients with prefrontal damage tested with free and forced report of general-knowledge answers, lesion mapping linked subcallosal ventromedial prefrontal cortex to impaired monitoring, while additional damage to posterior orbitofrontal cortex produced deficient control over which answers to volunteer (Hebscher et al., 2016).
- Across neurological patient studies, indices of frontal lobe integrity correlate strongly with metamemory accuracy, and frontal dysfunction combined with poor memory severely impairs both monitoring and control (Pannu & Kaszniak, 2005).
- In macaques making confidence judgments about their own recognition memory, reversible inactivation of prefrontal area 9/46d impaired metamemory for remote events and area 6 impaired it for recent events, without impairing recognition itself (Miyamoto et al., 2017).
Human Measurement
Level of Measurement: Directly Measured
- The standard report-option paradigm compares forced report with free report, in which people may withhold answers; free report raises accuracy substantially, usually at some cost in quantity, and both effects depend on the accuracy incentive and on how well confidence tracks correctness (Koriat & Goldsmith, 1996).
- A replication with 165 adult mock witnesses in Colombia found that using the report option raised the accuracy of reported answers by 12 percentage points (SD 9) among university students and by 7 points among adults with little formal education, so control worked in all three groups (Luna et al., 2023).
- In the study-time allocation paradigm, 74 self-paced learners recognized more words than yoked partners given the same total time (d_a 1.73 vs 1.51); the authors also cite a review in which 35 of 46 published conditions showed more study time given to items judged difficult (Tullis & Benjamin, 2011).
- Control over retrieval search was shown across five experiments: the longer people searched before answering "don't know", the higher the feeling of knowing that other participants reported for the same questions (Singer & Tiede, 2008).
AI Measurement
Level of Measurement: Directly Measured
AI Memory Performance: Below Human approximate provisional
- A peer-reviewed 2026 study tested this ability by name in 17 large language models (AI systems trained on vast amounts of text) using an answer-or-delegate game modelled on animal opt-out tasks, drawing on what the models learned in training; the link between being wrong and choosing to pass was weak (partial correlation 0.3 at best), and only one model, Claude 3.5 Sonnet, used passing to beat both its own and its teammate's accuracy (by 5.4 and 8.6 points) (Ackerman, 2026).
- A 2026 preprint found that four models (GPT-4o, Gemma 3 27B, DeepSeek, Qwen) decide to abstain on trivia questions using internal confidence and a threshold, as in the two-stage framework from animal metacognition; shifting confidence by directly altering the model's internal activity moved abstention from 66.5% to 7.0%, but no human data were collected (Kumaran et al., 2026).
- Across 20 frontier models and 20 datasets of unanswerable, underspecified or outdated questions, knowing when not to answer remained unsolved: larger models were of little help, and training for step-by-step reasoning reduced abstention by 24% on average (Kirichenko et al., 2025).
- For the current conversation and saved memory, withholding was tested with unanswerable trick questions about very long chats: scores (F1, out of 100) ranged from 2.5 (Claude 3 Sonnet) and 15.7 (GPT-4 Turbo) to 80.0 (Llama 3 70B, which could see only a short part of the chat), and from 9.9 to 32.5 when GPT-3.5 looked up stored chat excerpts; a human score of 89.4 is reported without saying how it was obtained, so it cannot be used (Maharana et al., 2024).
- The only documented human comparison found used the SelfAware set, which scores whether a model answers or declines unanswerable questions: GPT-4, the best of 20 models, reached F1 75.5, below both human volunteers (84.4 and 85.5), and other models scored lower, down to about 29 (Yin et al., 2023).
- This comparison is approximate: only two people each answered a random 100-question sample in 30 minutes, the samples may not match GPT-4's, and the questions have no answer at all rather than one the person failed to recall; it still points one way, since every model fell at least 8.9 F1 points below both people and the paper describes no outside help for either side (Yin et al., 2023).
- Same-task human–AI studies so far measure monitoring, not control: with 87 adults and GPT-4 on the same 60 quiz questions, confidence separated right from wrong answers less well for GPT-4 (type-2 AUROC 0.63) than for people on average (0.76, SD 0.14), though within the human spread, and no decision to answer or withhold was recorded (Yoshizawa et al., 2026).
Acerbi & Stubbersfield (2023)
Large language models show human-like content biases in transmission chain experiments
In
Proceedings of the National Academy of Sciences, 120(44)
Abstract
As the use of large language models (LLMs) grows, it is important to examine whether they exhibit biases in their output. Research in cultural evolution, using transmission chain experiments, demonstrates that humans have biases to attend to, remember, and transmit some types of content over others. Here, in five preregistered experiments using material from previous studies with human participants, we use the same, transmission chain-like methodology, and find that the LLM ChatGPT-3 shows biases analogous to humans for content that is gender-stereotype-consistent, social, negative, threat-related, and biologically counterintuitive, over other content. The presence of these biases in LLM...
Find it
Ackerman et al. (2005)
Working Memory and Intelligence: The Same or Different Constructs?
In
Psychological Bulletin, 131(1), 30-60
Find it
Ackerman (2026)
Evidence for Limited Metacognition in LLMs
In
arXiv
Abstract
The possibility of LLM self-awareness and even sentience is gaining increasing public attention and has major safety and policy implications, but the science of measuring them is still in a nascent state. Here we introduce a novel methodology for quantitatively evaluating metacognitive abilities in LLMs. Taking inspiration from research on metacognition in nonhuman animals, our approach eschews model self-reports and instead tests to what degree models can strategically deploy knowledge of internal states. Using two experimental paradigms, we demonstrate that frontier LLMs introduced since early 2024 show increasingly strong evidence of certain metacognitive abilities, specifically the...
Find it
Aerni et al. (2025)
Measuring Non-Adversarial Reproduction of Training Data in Large Language Models
In
arXiv
Abstract
Large language models memorize parts of their training data. Memorizing short snippets and facts is required to answer questions about the world and to be fluent in any language. But models have also been shown to reproduce long verbatim sequences of memorized text when prompted by a motivated adversary. In this work, we investigate an intermediate regime of memorization that we call non-adversarial reproduction, where we quantify the overlap between model responses and pretraining data when responding to natural and benign prompts. For a variety of innocuous prompt categories (e.g., writing a letter or a tutorial), we show that up to 15% of the text output by popular conversational...
Find it
Agren et al. (2012)
Disruption of Reconsolidation Erases a Fear Memory Trace in the Human Amygdala
In
Science, 337(6101), 1550-1552
Abstract
Memories become labile when recalled. In humans and rodents alike, reactivated fear memories can be attenuated by disrupting reconsolidation with extinction training. Using functional brain imaging, we found that, after a conditioned fear memory was formed, reactivation and reconsolidation left a memory trace in the basolateral amygdala that predicted subsequent fear expression and was tightly coupled to activity in the fear circuit of the brain. In contrast, reactivation followed by disrupted reconsolidation suppressed fear, abolished the memory trace, and attenuated fear-circuit connectivity. Thus, as previously demonstrated in rodents, fear memory suppression resulting from behavioral...
Find it
Akers et al. (2014)
Hippocampal Neurogenesis Regulates Forgetting During Adulthood and Infancy
In
Science, 344(6184), 598-602
Abstract
Forget It! When examining the relationship between the production of new neurons in the hippocampus and memory, studies have generally first manipulated hippocampal neurogenesis and afterward investigated memory formation and found that new neurons help to encode new memories. However, when investigating how similar manipulations of neurogenesis impact established hippocampus-dependent memories, Akers et al. (p. 598 ; see the Perspective by Mongiat and Schinder ) uncovered a role for neurogenesis in memory clearance. Thus, the continuous addition of new neurons both degrades existing information stored in hippocampal circuits and simultaneously provides substrates for new learning.
Find it
Alberts et al. (2024)
CURATe: Benchmarking Personalised Alignment of Conversational AI Assistants
In
arXiv
Abstract
We introduce a multi-turn benchmark for evaluating personalised alignment in LLM-based AI assistants, focusing on their ability to handle user-provided safety-critical contexts. Our assessment of ten leading models across five scenarios (with 337 use cases each) reveals systematic inconsistencies in maintaining user-specific consideration, with even top-rated "harmless" models making recommendations that should be recognised as obviously harmful to the user given the context provided. Key failure modes include inappropriate weighing of conflicting preferences, sycophancy (prioritising desires above safety), a lack of attentiveness to critical user information within the context window,...
Find it
Alfieri et al. (2013)
Learning Through Case Comparisons: A Meta-Analytic Review
In
Educational Psychologist, 48(2), 87-113
Find it
Anderson (1974)
Retrieval of propositional information from long-term memory
In
Cognitive Psychology, 6(4), 451-474
Find it
Anderson & Schooler (1991)
Reflections of the Environment in Memory
In
Psychological Science, 2(6), 396-408
Abstract
Availability of human memories for specific items shows reliable relationships to frequency, recency, and pattern of prior exposures to the item. These relationships have defied a systematic theoretical treatment. A number of environmental sources (New York Times, parental speech, electronic mail) are examined to show that the probability that a memory will be needed also shows reliable relationships to frequency, recency, and pattern of prior exposures. Moreover, the environmental relationships are the same as the memory relationships. It is argued that human memory has the form it does because it is adapted to these environmental relationships. Models for both the environment and human...
Find it
Anderson & Green (2001)
Suppressing unwanted memories by executive control
In
Nature, 410(6826), 366-369
Abstract
Freud proposed that unwanted memories can be forgotten by pushing them into the unconscious, a process called repression. The existence of repression has remained controversial for more than a century, in part because of its strong coupling with trauma, and the ethical and practical difficulties of studying such processes in controlled experiments. However, behavioural and neurobiological research on memory and attention shows that people have executive control processes directed at minimizing perceptual distraction, overcoming interference during short and long-term memory tasks and stopping strong habitual responses to stimuli. Here we show that these mechanisms can be recruited to...
Find it
Anderson & Hanslmayr (2014)
Neural mechanisms of motivated forgetting
In
Trends in Cognitive Sciences, 18(6), 279-292
Find it
Baddeley (2000)
The episodic buffer: a new component of working memory?
In
Trends in Cognitive Sciences, 4(11), 417-423
Find it
Baddeley (2012)
Working Memory: Theories, Models, and Controversies
In
Annual Review of Psychology, 63(1), 1-29
Abstract
I present an account of the origins and development of the multicomponent approach to working memory, making a distinction between the overall theoretical framework, which has remained relatively stable, and the attempts to build more specific models within this framework. I follow this with a brief discussion of alternative models and their relationship to the framework. I conclude with speculations on further developments and a comment on the value of attempting to apply models and theories beyond the laboratory studies on which they are typically based.
Find it
Balesni et al. (2024)
Lessons from Studying Two-Hop Latent Reasoning
In
arXiv
Abstract
Large language models can use chain-of-thought (CoT) to externalize reasoning, potentially enabling oversight of capable LLM agents. Prior work has shown that models struggle at two-hop question-answering without CoT. This capability is so basic that if it was a fundamental limitation, it would imply that many complex agentic tasks would similarly require CoT. We investigate LLM latent reasoning capabilities using two-hop question answering as a case study. Previous work on the gap between latent and externalized two-hop reasoning produced mixed evidence with inconclusive results. In this paper, we introduce a controlled setting for investigating two-hop reasoning in LLMs, where a...
Find it
Balleine & Dickinson (1998)
Goal-directed instrumental action: contingency and incentive learning and their cortical substrates
In
Neuropharmacology, 37(4-5), 407-419
Abstract
Instrumental behaviour is controlled by two systems: a stimulus–response habit mechanism and a goal-directed process that involves two forms of learning. The first is learning about the instrumental contingency between the response and reward, whereas the second consists of the acquisition of incentive value by the reward. Evidence for contingency learning comes from studies of reward devaluation and from demonstrations that instrumental performance is sensitive not only the probability of contiguous reward but also to the probability of unpaired rewards. The process of incentive learning is evident in the acquisition of control over performance by primary motivational states. Preliminary...
Find it
Balleine & O'Doherty (2010)
Human and Rodent Homologies in Action Control: Corticostriatal Determinants of Goal-Directed and Habitual Action
In
Neuropsychopharmacology, 35(1), 48-69
Find it
Barbey et al. (2013)
Dorsolateral prefrontal contributions to human working memory
In
Cortex, 49(5), 1195-1205
Find it
Barthel et al. (2025)
Habituation of the biological response to repeated psychosocial stress: a systematic review and meta-analysis
In
Neuroscience & Biobehavioral Reviews, 169, 105996
Find it
Bartlett (1932)
Remembering: A Study in Experimental and Social Psychology
In
Cambridge University Press
Find it
Barwood et al. (2024)
Habituation of the cold shock response: A systematic review and meta-analysis
In
Journal of Thermal Biology, 119, 103775
Find it
Barzykowski & Niedźwieńska (2018)
Involuntary autobiographical memories are relatively more often reported during high cognitive load tasks
In
Acta Psychologica, 182, 119-128
Find it
Batchelder & Riefer (1990)
Multinomial processing models of source monitoring.
In
Psychological Review, 97(4), 548-564
Find it
Beck & Weber (2024)
Who said it? Native and non-native listeners' source memory for object-speaker associations
In
Frontiers in Language Sciences, 2
Abstract
Memory for who said what can help native and non-native listeners identify critical information for use in conversations. In two experiments, source memory for object-speaker associations was tested for listeners varying in language proficiency. Native and non-native participants of English first heard speakers name objects with prototypical color (e.g., broccoli) before they had to indicate who had previously named each object. Speakers either consistently named objects of one color, or they randomly named objects of different colors. While both listener groups associated objects with speakers successfully, in Experiment 1, with black and white images, native listeners showed an...
Find it
Benoit & Anderson (2012)
Opposing Mechanisms Support the Voluntary Forgetting of Unwanted Memories
In
Neuron, 76(2), 450-460
Abstract
Reminders of the past can trigger the recollection of events that one would rather forget. Here, using fMRI, we demonstrate two distinct neural mechanisms that foster the intentional forgetting of such unwanted memories. Both mechanisms impair long-term retention by limiting momentary awareness of the memories, yet they operate in opposite ways. One mechanism, direct suppression, disengages episodic retrieval through the systemic inhibition of hippocampal processing that originates from right dorsolateral prefrontal cortex (PFC). The opposite mechanism, thought substitution, instead engages retrieval processes to occupy the limited focus of awareness with a substitute memory. It is...
Find it
Berglund et al. (2023)
The Reversal Curse: LLMs trained on "A is B" fail to learn "B is A"
In
arXiv
Abstract
We expose a surprising failure of generalization in auto-regressive large language models (LLMs). If a model is trained on a sentence of the form "A is B", it will not automatically generalize to the reverse direction "B is A". This is the Reversal Curse. For instance, if a model is trained on "Valentina Tereshkova was the first woman to travel to space", it will not automatically be able to answer the question, "Who was the first woman to travel to space?". Moreover, the likelihood of the correct answer ("Valentina Tershkova") will not be higher than for a random name. Thus, models do not generalize a prevalent pattern in their training set: if "A is B" occurs, "B is A" is more likely to...
Find it
Bergman & Roediger (1999)
Can Bartlett’s repeated reproduction experiments be replicated?
In
Memory & Cognition, 27(6), 937-947
Find it
Berntsen et al. (2013)
Why am I remembering this now? Predicting the occurrence of involuntary (spontaneous) episodic memories.
In
Journal of Experimental Psychology: General, 142(2), 426-444
Abstract
Involuntary episodic memories are memories of events that come to mind spontaneously, that is, with no preceding retrieval attempts. They are common in daily life and observed in a range of clinical disorders in the form of negative, intrusive recollections or flashbacks. However, little is known about their underlying mechanisms. Here we report a series of experiments in which—for the first time—the activation of involuntary memories is controlled and predicted on the basis of manipulations done at encoding. During encoding, participants were presented with pictures of scenes paired with sounds. Both scene and sound could be either unique (derived from a category that was presented only...
Find it
Berntsen et al. (2017)
Spontaneous or intentional? Involuntary versus voluntary episodic memories in older and younger adults.
In
Psychology and Aging, 32(2), 192-201
Find it
Berntsen (2021)
Involuntary autobiographical memories and their relation to other forms of spontaneous thoughts
In
Philosophical Transactions of the Royal Society B: Biological Sciences, 376(1817), 20190693
Abstract
Involuntary autobiographical memories are memories of personal events that come to mind spontaneously—that is, with no conscious initiation of the retrieval process. Such spontaneously arising memories were long ignored in cognitive psychology, which generally has focused on controlled and strategic forms of remembering, studied in laboratory settings. Recent evidence shows that involuntary memories of past events are highly frequent in daily life, and that they represent a context-sensitive, and associative way of recollecting past events that involves little executive control. They operate by constraints that favour recent events and events with a distinct feature overlap to the current...
Find it
Berres & Erdfelder (2021)
The sleep benefit in episodic memory: An integrative review and a meta-analysis.
In
Psychological Bulletin, 147(12), 1309-1353
Abstract
People recall more information after sleep than after an equally long period of wakefulness. This sleep benefit in episodic memory has been documented in almost a century of research. However, an integrative review of hypothesized underlying processes, a comprehensive quantification of the benefit, and a systematic investigation of potential moderators has been missing so far. Here, we address these issues by analyzing 823 effect sizes from 271 independent samples that were reported in 177 articles published between 1967 and 2019. Using multilevel metaregressions with robust variance estimates, we found a moderate overall sleep benefit in episodic memory (g = 0.44). Moderator analyses...
Find it
Binder et al. (2009)
Where Is the Semantic System? A Critical Review and Meta-Analysis of 120 Functional Neuroimaging Studies
In
Cerebral Cortex, 19(12), 2767-2796
Abstract
Semantic memory refers to knowledge about people, objects, actions, relations, self, and culture acquired through experience. The neural systems that store and retrieve this information have been studied for many years, but a consensus regarding their identity has not been reached. Using strict inclusion criteria, we analyzed 120 functional neuroimaging studies focusing on semantic processing. Reliable areas of activation in these studies were identified using the activation likelihood estimate (ALE) technique. These activations formed a distinct, left-lateralized network comprised of 7 regions: posterior inferior parietal lobe, middle temporal gyrus, fusiform and parahippocampal gyri,...
Find it
Binz & Schulz (2023)
Using cognitive psychology to understand GPT-3
In
Proceedings of the National Academy of Sciences, 120(6)
Abstract
We study GPT-3, a recent large language model, using tools from cognitive psychology. More specifically, we assess GPT-3's decision-making, information search, deliberation, and causal reasoning abilities on a battery of canonical experiments from the literature. We find that much of GPT-3's behavior is impressive: It solves vignette-based tasks similarly or better than human subjects, is able to make decent decisions from descriptions, outperforms humans in a multiarmed bandit task, and shows signatures of model-based reinforcement learning. Yet, we also find that small perturbations to vignette-based tasks can lead GPT-3 vastly astray, that it shows no signatures of directed...
Find it
Block et al. (2025)
Model-Based Person Fit Statistics Applied to the Wechsler Adult Intelligence Scale IV
In
Educational and Psychological Measurement, 85(6), 1123-1151
Abstract
An important task in clinical neuropsychology is to evaluate whether scores obtained on a test battery, such as the Wechsler Adult Intelligence Scale Fourth Edition (WAIS-IV), can be considered “credible” or “valid” for a particular patient. Such evaluations are typically made based on responses to performance validity tests (PVTs). As a complement to PVTs, we propose that WAIS-IV profiles also be evaluated using a residual-based M-distance ( d ri 2 ) person fit statistic. Large d ri 2 values flag profiles that are inconsistent with the factor analytic model underlying the interpretation of test scores. We first established a well-fitting model with four correlated factors for 10 core...
Find it
Born & Wilhelm (2012)
System consolidation of memory during sleep
In
Psychological Research, 76(2), 192-203
Find it
Bovet et al. (2024)
Direct and indirect linguistic measures of common ground in dialogue studies involving a matching task: A systematic review
In
Psychonomic Bulletin & Review, 31(1), 122-136
Abstract
Abstract During dialogue, speakers attempt to adapt messages to their addressee appropriately by taking into consideration their common ground (i.e., all the information mutually known by the conversational partners) to ensure successful communication. Knowing and remembering what information is part of the common ground shared with a given partner and using it during dialogue are crucial skills for social interaction. It is therefore important to better understand how we can measure the use of common ground and to identify the potential associated psychological processes. In this context, a systematic review of the literature was performed to list the linguistic measures of common ground...
Find it
Bowles et al. (2007)
Impaired familiarity with preserved recollection after anterior temporal-lobe resection that spares the hippocampus
In
Proceedings of the National Academy of Sciences, 104(41), 16382-16387
Abstract
It is well established that the medial-temporal lobe (MTL) is critical for recognition memory. The MTL is known to be composed of distinct structures that are organized in a hierarchical manner. At present, it remains controversial whether lower structures in this hierarchy, such as perirhinal cortex, support memory functions that are distinct from those of higher structures, in particular the hippocampus. Perirhinal cortex has been proposed to play a specific role in the assessment of familiarity during recognition, which can be distinguished from the selective contributions of the hippocampus to the recollection of episodic detail. Some researchers have argued, however, that the...
Find it
Bowman & Zeithamova (2018)
Abstract Memory Representations in the Ventromedial Prefrontal Cortex and Hippocampus Support Concept Generalization
In
The Journal of Neuroscience, 38(10), 2605-2614
Abstract
Memory function involves both the ability to remember details of individual experiences and the ability to link information across events to create new knowledge. Prior research has identified the ventromedial prefrontal cortex (VMPFC) and the hippocampus as important for integrating across events in the service of generalization in episodic memory. The degree to which these memory integration mechanisms contribute to other forms of generalization, such as concept learning, is unclear. The present study used a concept-learning task in humans (both sexes) coupled with model-based fMRI to test whether VMPFC and hippocampus contribute to concept generalization, and whether they do so by...
Find it
Brady et al. (2008)
Visual long-term memory has a massive storage capacity for object details
In
Proceedings of the National Academy of Sciences, 105(38), 14325-14329
Abstract
One of the major lessons of memory research has been that human memory is fallible, imprecise, and subject to interference. Thus, although observers can remember thousands of images, it is widely assumed that these memories lack detail. Contrary to this assumption, here we show that long-term memory is capable of storing a massive number of objects with details from the image. Participants viewed pictures of 2,500 objects over the course of 5.5 h. Afterward, they were shown pairs of images and indicated which of the two they had seen. The previously viewed item could be paired with either an object from a novel category, an object of the same basic-level category, or the same object in a...
Find it
Breiter et al. (1996)
Response and Habituation of the Human Amygdala during Visual Processing of Facial Expression
In
Neuron, 17(5), 875-887
Find it
Brennan & Clark (1996)
Conceptual pacts and lexical choice in conversation.
In
Journal of Experimental Psychology: Learning, Memory, and Cognition, 22(6), 1482-1493
Find it
Brewer & Wells (2006)
The confidence-accuracy relationship in eyewitness identification: Effects of lineup instructions, foil similarity, and target-absent base rates.
In
Journal of Experimental Psychology: Applied, 12(1), 11-30
Find it
Brookwell et al. (2013)
Externalizing biases and hallucinations in source-monitoring, self-monitoring and signal detection studies: a meta-analytic review
In
Psychological Medicine, 43(12), 2465-2475
Abstract
Background Cognitive models have postulated that auditory hallucinations arise from the misattribution of internally generated cognitive events to external sources. Several experimental paradigms have been developed to assess this externalizing bias in clinical and non-clinical hallucination-prone samples, including source-monitoring, verbal self-monitoring and auditory signal detection tasks. This meta-analysis aims to synthesize the wealth of empirical findings from these experimental studies. Method A database search was carried out for reports between January 1985 and March 2012. Additional studies were retrieved by contacting authors and screening references of eligible reports....
Find it
Brown-Schmidt (2009)
Partner-specific interpretation of maintained referential precedents during interactive dialog
In
Journal of Memory and Language, 61(2), 171-190
Find it
Buda et al. (2011)
A Specific Brain Structural Basis for Individual Differences in Reality Monitoring
In
The Journal of Neuroscience, 31(40), 14308-14313
Abstract
Much recent interest has centered on understanding the relationship between brain structure variability and individual differences in cognition, but there has been little progress in identifying specific neuroanatomical bases of such individual differences. One cognitive ability that exhibits considerable variability in the healthy population is reality monitoring; the cognitive processes used to introspectively judge whether a memory came from an internal or external source (e.g., whether an event was imagined or actually occurred). Neuroimaging research has implicated the medial anterior prefrontal cortex (PFC) in reality monitoring, and here we sought to determine whether morphological...
Find it
Bunsey & Eichenbaum (1996)
Conservation of hippocampal memory function in rats and humans
In
Nature, 379(6562), 255-257
Find it
Burgess et al. (2011)
Functional neuroimaging studies of prospective memory: What have we learnt so far?
In
Neuropsychologia, 49(8), 2246-2257
Abstract
The complexity of the behaviour described by the term “prospective memory” meant that it was not at all clear, when the earliest studies were conducted, that this would prove a fruitful area for neuroimaging study. However, a consistent relation rapidly emerged between activation in rostral prefrontal cortex (approximating Brodmann Area 10) and performance of prospective memory paradigms. This consistency has greatly increased the accumulation of findings, since each study has offered perspectives on the previous ones. Considerable help too has come from broad agreement between functional neuroimaging findings and those from other methods (e.g. human lesion studies, electrophysiology)....
Find it
Byrne & Hawkins (2015)
Nonassociative Learning in Invertebrates
In
Cold Spring Harbor Perspectives in Biology, 7(5), a021675
Find it
Cai et al. (2024)
Do large language models resemble humans in language use?
In
Proceedings of the Workshop on Cognitive Modeling and Computational Linguistics, 37-56
Find it
Cao et al. (2025)
Analyzing Memory Effects in Large Language Models through the lens of Cognitive Psychology
In
arXiv
Abstract
Memory, a fundamental component of human cognition, exhibits adaptive yet fallible characteristics as illustrated by Schacter's memory "sins".These cognitive phenomena have been studied extensively in psychology and neuroscience, but the extent to which artificial systems, specifically Large Language Models (LLMs), emulate these cognitive phenomena remains underexplored. This study uses human memory research as a lens for understanding LLMs and systematically investigates human memory effects in state-of-the-art LLMs using paradigms drawn from psychological research. We evaluate seven key memory phenomena, comparing human behavior to LLM performance. Both people and models remember less...
Find it
Carini et al. (2025)
Exploring the Link Between the Emotional Recall Task and Mental Health in Humans and LLMs
In
Information, 16(12), 1057
Abstract
The ability of large language models to recall human emotions provides a novel opportunity to investigate links among memory, affect, and mental health. This study explores whether the Emotional Recall Task (ERT), a free word-association paradigm, can reveal cognitive markers of distress in both humans and large language models (LLMs). Using spreading activation simulations grounded in cognitive network science, we examined how the recall of emotional concepts (e.g., stress, anxiety, and depression) relates to psychometric measures of well-being and personality. In Study 1, correlations were tested between activation dynamics and clinical scales (DASS-21, PANAS, and Life Satisfaction) in...
Find it
Carpenter & Schacter (2017)
Flexible retrieval: When true inferences produce false memories.
In
Journal of Experimental Psychology: Learning, Memory, and Cognition, 43(3), 335-349
Find it
Cash et al. (2025)
Quantifying uncert-AI-nty: Testing the accuracy of LLMs’ confidence judgments
In
Memory & Cognition, 54(2), 375-400
Abstract
Abstract The rise of Large Language Model (LLM) chatbots, such as ChatGPT and Gemini, has revolutionized how we access information. These LLMs can answer a wide array of questions on nearly any topic. When humans answer questions, especially difficult or uncertain questions, they often accompany their responses with metacognitive confidence judgments indicating their belief in their accuracy. LLMs are certainly capable of providing confidence judgments, but it is currently unclear how accurate these confidence judgments are. To fill this gap in the literature, the present studies investigate the capability of LLMs to quantify uncertainty through confidence judgments. We compare the...
Find it
Castellucci et al. (1970)
Neuronal Mechanisms of Habituation and Dishabituation of the Gill-Withdrawal Reflex in Aplysia
In
Science, 167(3926), 1745-1748
Abstract
The cellular mechanisms of habituation and dishabituation of the gill-withdrawal reflex in Aplysia were studied with an isolated abdominal ganglion connected to a piece of skin from the tactile receptive field of the reflex. By obtaining simultaneous intracellular recordings from both the sensory neurons and one of the main identified motor neurons, we have been able to reduce the reflex to its monosynaptic components. The monosynaptic excitatory postsynaptic potentials showed a profound low-frequency depression when repeatedly elicited and showed heterosynaptic facilitation after application of a strong stimulus to another pathway. Thus, both habituation and dishabituation can be...
Find it
Cepeda et al. (2006)
Distributed practice in verbal recall tasks: A review and quantitative synthesis.
In
Psychological Bulletin, 132(3), 354-380
Find it
Cepeda et al. (2008)
Spacing Effects in Learning
In
Psychological Science, 19(11), 1095-1102
Abstract
To achieve enduring retention, people must usually study information on multiple occasions. How does the timing of study events affect retention? Prior research has examined this issue only in a spotty fashion, usually with very short time intervals. In a study aimed at characterizing spacing effects over significant durations, more than 1,350 individuals were taught a set of facts and—after a gap of up to 3.5 months—given a review. A final test was administered at a further delay of up to 1 year. At any given test delay, an increase in the interstudy gap at first increased, and then gradually reduced, final test performance. The optimal gap increased as test delay increased. However,...
Find it
Chalkia et al. (2020)
No persistent attenuation of fear memories in humans: A registered replication of the reactivation-extinction effect
In
Cortex, 129, 496-509
Find it
Chang et al. (2024)
How Do Large Language Models Acquire Factual Knowledge During Pretraining?
In
arXiv
Abstract
Despite the recent observation that large language models (LLMs) can store substantial factual knowledge, there is a limited understanding of the mechanisms of how they acquire factual knowledge through pretraining. This work addresses this gap by studying how LLMs acquire factual knowledge during pretraining. The findings reveal several important insights into the dynamics of factual knowledge acquisition during pretraining. First, counterintuitively, we observe that pretraining on more data shows no significant improvement in the model's capability to acquire and maintain factual knowledge. Next, there is a power-law relationship between training steps and forgetting of memorization and...
Find it
Charles & Bowman (2025)
Age-related differences in associative inference are larger than differences in direct associative memory.
In
Psychology and Aging, 40(5), 524-536
Find it
Chattaraj & Raj (2026)
LLMs Remember First, Forget Last: Dual-Process Interference in Large Language Models
In
arXiv
Abstract
Large language models can process millions of tokens, yet how they handle conflicting information within context remains poorly understood. From patient health logs tracking evolving vital signs to legal documents with superseding clauses, real-world applications routinely require models to retrieve specific values from streams of semantically similar, competing updates. We adapt classical interference paradigms from cognitive psychology to compare retroactive interference (RI; recalling initial values after updates) and proactive interference (PI; recalling recent values despite competing prior encodings) across 39 LLMs spanning 1B to 2.5T parameters. Every model exhibits the same...
Find it
Chauvet (2024)
Memory Traces: Are Transformers Tulving Machines?
In
arXiv
Abstract
Memory traces--changes in the memory system that result from the perception and encoding of an event--were measured in pioneering studies by Endel Tulving and Michael J. Watkins in 1975. These and further experiments informed the maturation of Tulving's memory model, from the GAPS (General Abstract Processing System} to the SPI (Serial-Parallel Independent) model. Having current top of the line LLMs revisit the original Tulving-Watkins tests may help in assessing whether foundation models completely instantiate or not this class of psychological models.
Find it
Chauvet (2024)
Memory GAPS: Would LLMs pass the Tulving Test?
In
arXiv
Abstract
The Tulving Test was designed to investigate memory performance in recognition and recall tasks. Its results help assess the relevance of the "Synergistic Ecphory Model" of memory and similar RK paradigms in human performance. This paper starts investigating whether the more than forty-year-old framework sheds some light on LLMs' acts of remembering.
Find it
Cheng et al. (2008)
Neural substrates underlying human delay and trace eyeblink conditioning
In
Proceedings of the National Academy of Sciences, 105(23), 8108-8113
Abstract
Classical conditioning paradigms, such as trace conditioning, in which a silent period elapses between the offset of the conditioned stimulus (CS) and the delivery of the unconditioned stimulus (US), and delay conditioning, in which the CS and US coterminate, are widely used to study the neural substrates of associative learning. However, there are significant gaps in our knowledge of the neural systems underlying conditioning in humans. For example, evidence from animal and human patient research suggests that the hippocampus plays a critical role during trace eyeblink conditioning, but there is no evidence to date in humans that the hippocampus is active during trace eyeblink...
Find it
Cheung (2026)
Kindling in Neural Systems: Progressive Adversarial Sensitization During LLM Alignment Mirrors Psychiatric Progression
Abstract
Abstract Objective Reinforcement learning from human feedback (RLHF) is widely used to make large language models safer, yet repeated preference tuning could also make them easier to breach. Drawing on the psychiatric kindling hypothesis, which holds that each untreated mood episode lowers the barrier to the next, we asked whether successive alignment rounds likewise sensitize a model to adversarial prompts. Methods A 1.1-billion-parameter chat model (TinyLlama−1.1B-Chat) equipped with LoRA adapters completed ten preference-tuning cycles. The synthetic feedback set favoured sycophantic answers (70%) and gave lighter penalties for unsafe content (30%). Three experimental arms were...
Find it
Chhikara et al. (2025)
Mem0: Building Production-Ready AI Agents with Scalable Long-Term Memory
In
Frontiers in Artificial Intelligence and Applications
Abstract
Large Language Models (LLMs) have demonstrated remarkable prowess in generating contextually coherent responses, yet their fixed context windows pose fundamental challenges for maintaining consistency over prolonged multi-session dialogues. We introduce Mem0, a scalable memory-centric architecture that addresses this issue by dynamically extracting, consolidating, and retrieving salient information from ongoing conversations. Building on this foundation, we further propose an enhanced variant that leverages graph-based memory representations to capture complex relational structures among conversational elements. Through comprehensive evaluations on LOCOMO benchmark, we systematically...
Find it
Christian & Thompson (2003)
Neural Substrates of Eyeblink Conditioning: Acquisition and Retention
In
Learning & Memory, 10(6), 427-455
Abstract
Classical conditioning of the eyeblink reflex to a neutral stimulus that predicts an aversive stimulus is a basic form of associative learning. Acquisition and retention of this learned response require the cerebellum and associated sensory and motor pathways and engage several other brain regions including the hippocampus, neocortex, neostriatum, septum, and amygdala. The cerebellum and its associated circuitry form the essential neural system for delay eyeblink conditioning. Trace eyeblink conditioning, a learning paradigm in which the conditioned and unconditioned stimuli are noncontiguous, requires both the cerebellum and the hippocampus and exhibits striking parallels to declarative...
Find it
Christophel et al. (2017)
The Distributed Nature of Working Memory
In
Trends in Cognitive Sciences, 21(2), 111-124
Find it
Clark & Squire (1998)
Classical Conditioning and Brain Systems: The Role of Awareness
In
Science, 280(5360), 77-81
Abstract
Classical conditioning of the eye-blink response, perhaps the best studied example of associative learning in vertebrates, is relatively automatic and reflexive, and with the standard procedure (simple delay conditioning), it is intact in animals with hippocampal lesions. In delay conditioning, a tone \[the conditioned stimulus (CS)\] is presented just before an air puff to the eye \[the unconditioned stimulus (US)\]. The US is then presented, and the two stimuli coterminate. In trace conditioning, a variant of the standard paradigm, a short interval (500 to 1000 ms) is interposed between the offset of the CS and the onset of the US. Animals with hippocampal lesions fail to acquire trace...
Find it
Clark et al. (2002)
Classical conditioning, awareness, and brain systems
In
Trends in Cognitive Sciences, 6(12), 524-531
Find it
Clark et al. (2016)
Intrusive memories to traumatic footage: the neural basis of their encoding and involuntary recall
In
Psychological Medicine, 46(3), 505-518
Abstract
BackgroundA hallmark symptom after psychological trauma is the presence of intrusive memories. It is unclear why only some moments of trauma become intrusive, and how these memories involuntarily return to mind. Understanding the neural mechanisms involved in the encoding and involuntary recall of intrusive memories may elucidate these questions.MethodParticipants (n= 35) underwent functional magnetic resonance imaging (fMRI) while being exposed to traumatic film footage. After film viewing, participants indicated within the scanner, while undergoing fMRI, if they experienced an intrusive memory of the film. Further intrusive memories in daily life were recorded for 7 days. After 7 days,...
Find it
Clark et al. (2026)
Forgetting as a consequence of retrieval suppression: A meta-analytic review.
In
Journal of Experimental Psychology: General, 155(5), 1318-1346
Abstract
Suppressing retrieval of a memory when faced with a reminder has been shown to reduce the probability of that memory coming to mind in the future, a phenomenon known as suppression-induced forgetting (SIF). However, efforts to observe SIF have sometimes failed, leading some to question the robustness of the effect. To address this issue, we conducted a comprehensive meta-analysis of SIF within the think/no-think paradigm including several key methodological variants to the original design. Across 500 effects from 120 studies, we observed a consistent aggregate SIF effect (d ≈ 0.20–0.40) measured using either same probe (SIF ≈ 6%), independent probe (SIF ≈ 5%–6%), or recognition (SIF ≈...
Find it
Clayton & Dickinson (1998)
Episodic-like memory during cache recovery by scrub jays
In
Nature, 395(6699), 272-274
Find it
Coda-Forno et al. (2024)
CogBench: a large language model walks into a psychology lab
In
arXiv
Abstract
Large language models (LLMs) have significantly advanced the field of artificial intelligence. Yet, evaluating them comprehensively remains challenging. We argue that this is partly due to the predominant focus on performance metrics in most benchmarks. This paper introduces CogBench, a benchmark that includes ten behavioral metrics derived from seven cognitive psychology experiments. This novel approach offers a toolkit for phenotyping LLMs' behavior. We apply CogBench to 35 LLMs, yielding a rich and diverse dataset. We analyze this data using statistical multilevel modeling techniques, accounting for the nested dependencies among fine-tuned versions of specific LLMs. Our study...
Find it
Cohen & Squire (1980)
Preserved Learning and Retention of Pattern-Analyzing Skill in Amnesia: Dissociation of Knowing How and Knowing That
In
Science, 210(4466), 207-210
Abstract
Amnesic patients acquired a mirror-reading skill at a rate equivalent to that of matched control subjects and retained it for at least 3 months. The results indicate that the class of preserved learning skills in amnesia is broader than previously reported. Amnesia seems to spare information that is based on rules or procedures, as contrasted with information that is data-based or declarative—"knowing how" rather than "knowing that." The results support the hypothesis that such a distinction is honored by the nervous system.
Find it
Cona et al. (2015)
Neural bases of prospective memory: A meta-analysis and the “Attention to Delayed Intention” (AtoDI) model
In
Neuroscience & Biobehavioral Reviews, 52, 21-37
Find it
Cona et al. (2016)
Effects of cue focality on the neural mechanisms of prospective memory: A meta-analysis of neuroimaging studies
In
Scientific Reports, 6(1)
Abstract
AbstractRemembering to execute pre-defined intentions at the appropriate time in the future is typically referred to as Prospective Memory (PM). Studies of PM showed that distinct cognitive processes underlie the execution of delayed intentions depending on whether the cue associated with such intentions is focal to ongoing activity processing or not (i.e., cue focality). The present activation likelihood estimation (ALE) meta-analysis revealed several differences in brain activity as a function of focality of the PM cue. The retrieval of intention is supported mainly by left anterior prefrontal cortex (Brodmann Area, BA 10) in nonfocal tasks, and by cerebellum and ventral parietal...
Find it
Conway & Pleydell-Pearce (2000)
The construction of autobiographical memories in the self-memory system.
In
Psychological Review, 107(2), 261-288
Abstract
The authors describe a model of autobiographical memory in which memories are transitory mental constructions within a self-memory system (SMS). The SMS contains an autobiographical knowledge base and current goals of the working self. Within the SMS, control processes modulate access to the knowledge base by successively shaping cues used to activate autobiographical memory knowledge structures and, in this way, form specific memories. The relation of the knowledge base to active goals is reciprocal, and the knowledge base "grounds" the goals of the working self. It is shown how this model can be used to draw together a wide range of diverse data from cognitive, social, developmental,...
Find it
Cordi & Rasch (2021)
How robust are sleep-mediated memory benefits?
In
Current Opinion in Neurobiology, 67, 1-7
Find it
Cowan (2001)
The magical number 4 in short-term memory: A reconsideration of mental storage capacity
In
Behavioral and Brain Sciences, 24(1), 87-114
Abstract
Miller (1956) summarized evidence that people can remember about seven chunks in short-term memory (STM) tasks. However, that number was meant more as a rough estimate and a rhetorical device than as a real capacity limit. Others have since suggested that there is a more precise capacity limit, but that it is only three to five chunks. The present target article brings together a wide variety of data on capacity limits suggesting that the smaller capacity limit is real. Capacity limits will be useful in analyses of information processing only if the boundary conditions for observing them can be carefully described. Four basic conditions in which chunks can be identified and capacity...
Find it
Damiani et al. (2022)
Understanding source monitoring subtypes and their relation to psychosis: a systematic review and meta‐analysis
In
Psychiatry and Clinical Neurosciences, 76(5), 162-171
Abstract
**Aims:** Source monitoring (SM) is the metacognitive ability to determine the origin of one's experiences. SM is altered in primary psychiatric psychosis, although relationships between SM subtypes, other cognitive domains and symptoms are unclear. Our aims were to synthesize evidence comparing psychosis -with and without hallucinations- and healthy controls classifying SM subtypes by source discrimination (internal/external/reality monitoring) and stimulus modality (visual/auditory/imagined/performed). **Methods:** This systematic review adopted Preferred Reporting Items for Systematic Reviews and Meta-Analyses, Meta-analyses Of Observational Studies in Epidemiology and Population,...
Find it
Damiani et al. (2024)
The noise in our brain: A systematic review and meta-analysis of neuroimaging and signal-detection studies on source monitoring in psychosis
In
Journal of Psychiatric Research, 169, 142-151
Find it
Davachi et al. (2003)
Multiple routes to memory: Distinct medial temporal lobe processes build item and source memories
In
Proceedings of the National Academy of Sciences, 100(4), 2157-2162
Abstract
A central function of memory is to permit an organism to distinguish between stimuli that have been previously encountered and those that are novel. Although the medial temporal lobe (which includes the hippocampus and surrounding perirhinal, parahippocampal, and entorhinal cortices) is known to be crucial for recognition memory, controversy remains regarding how the specific subregions within the medial temporal lobe contribute to recognition. We used event-related functional MRI to examine the relation between activation in distinct medial temporal lobe subregions during memory formation and the ability ( i ) to later recognize an item as previously encountered (item recognition) and (...
Find it
Daviddi et al. (2023)
Common and distinct correlates of construction and elaboration of episodic-autobiographical memory: An ALE meta-analysis
In
Cortex, 163, 123-138
Find it
Davis & Zhong (2017)
The Biology of Forgetting—A Perspective
In
Neuron, 95(3), 490-503
Find it
Daw et al. (2011)
Model-Based Influences on Humans' Choices and Striatal Prediction Errors
In
Neuron, 69(6), 1204-1215
Abstract
The mesostriatal dopamine system is prominently implicated in model-free reinforcement learning, with fMRI BOLD signals in ventral striatum notably covarying with model-free prediction errors. However, latent learning and devaluation studies show that behavior also shows hallmarks of model-based planning, and the interaction between model-based and model-free values, prediction errors, and preferences is underexplored. We designed a multistep decision task in which model-based and model-free influences on human choice behavior could be distinguished. By showing that choices reflected both influences we could then test the purity of the ventral striatal BOLD signal as a model-free report....
Find it
de Wit et al. (2018)
Shifting the balance between goals and habits: Five failures in experimental habit induction.
In
Journal of Experimental Psychology: General, 147(7), 1043-1065
Abstract
Habits are repetitive behaviors that become ingrained with practice, routine, and repetition. The more we repeat an action, the stronger our habits become. Behavioral and clinical neuroscientists have become increasingly interested in this topic because habits may contribute to aspects of maladaptive human behavior, such as compulsive behavior in psychiatry. Numerous studies have demonstrated that habits can be induced in otherwise healthy rats by simply overtraining stimulus–response behaviors. However, despite growing interest in this topic and its application to psychiatry, a similar body of work in humans is absent. Only a single study has been published in humans that shows the...
Find it
Debiec et al. (2002)
Cellular and Systems Reconsolidation in the Hippocampus
In
Neuron, 36(3), 527-538
Find it
Dede et al. (2013)
Hippocampal damage impairs recognition memory broadly, affecting both parameters in two prominent models of memory
In
Proceedings of the National Academy of Sciences, 110(16), 6577-6582
Abstract
Declarative memory is thought to rely on two processes: recollection and familiarity. Recollection involves remembering specific details about the episode in which an item was encountered, and familiarity involves simply knowing that an item was presented even when no information can be recalled about the episode itself. There has been debate whether the hippocampus supports only recollection or whether it supports both processes. We approached this issue in a relatively theory-neutral way by fitting two prominent models that have been used to describe recognition memory: dual process signal detection and unequal variance signal detection. Both models yield two parameters of interest when...
Find it
Deeb & Roger (2024)
Do Unlearning Methods Remove Information from Language Model Weights?
In
arXiv
Abstract
Large Language Models' knowledge of how to perform cyber-security attacks, create bioweapons, and manipulate humans poses risks of misuse. Previous work has proposed methods to unlearn this knowledge. Historically, it has been unclear whether unlearning techniques are removing information from the model weights or just making it harder to access. To disentangle these two objectives, we propose an adversarial evaluation method to test for the removal of information from model weights: we give an attacker access to some facts that were supposed to be removed, and using those, the attacker tries to recover other facts from the same distribution that cannot be guessed from the accessible...
Find it
Dennis et al. (2026)
Beyond Inference-Only Deployment: Comparing Weight-Based Consolidation Against Cascading Compaction
In
arXiv
Abstract
Major LLM platforms deploy models in an inference-only configuration: the model serves requests but never updates per-user weights. Users must repeatedly re-teach preferences, corrections, and project context, and context-based workarounds consume context-window space and degrade under cascading compaction. We evaluate an alternative: nightly consolidation of interaction knowledge into model weights via reflection, synthesis, and Low-Rank Adaptation (LoRA) fine-tuning on a single consumer GPU. Across ten realistic software development conversations (n = 10, 1,146 test questions across three memory types), three cycles of cascading compaction retain 36.8 +/- 3.0% of knowledge (between an...
Find it
Devaluez et al. (2023)
Episodic and semantic feeling-of-knowing in aging: a systematic review and meta-analysis
In
Scientific Reports, 13(1)
Abstract
AbstractA complex pattern of preservation and deterioration in metacognition in aging is found, especially regarding predicting future memory retrieval (i.e., feeling-of-knowing, FOK). While semantic FOK (sFOK) is preserved with age, studies on episodic tasks (eFOK) produce equivocal findings. We present a meta-analysis of 20 studies on eFOK and sFOK, analyzing the difference in metacognitive sensitivity between 922 younger and 966 older adults, taking into account the difference in memory performance. The sFOK studies yielded no overall age effect (8 effects, g = −0.10 [−0.29, 0.10]). However, we found a reliable age-group difference on eFOK (22 effects, g = 0.53 [0.28, 0.78]), which was...
Find it
Diana et al. (2007)
Imaging recollection and familiarity in the medial temporal lobe: a three-component model
In
Trends in Cognitive Sciences, 11(9), 379-386
Abstract
The medial temporal lobe (MTL) plays a crucial role in supporting memory for events, but the functional organization of regions in the MTL remains controversial, especially regarding the extent to which different subregions support recognition based on familiarity or recollection. Here we review results from functional neuroimaging studies showing that, whereas activity in the hippocampus and posterior parahippocampal gyrus is disproportionately associated with recollection, activity in the anterior parahippocampal gyrus is disproportionately associated with familiarity. The results are consistent with the idea that the parahippocampal cortex (located in the posterior parahippocampal...
Find it
Dickinson (1985)
Actions and habits: the development of behavioural autonomy
In
Philosophical Transactions of the Royal Society of London. B, Biological Sciences, 308(1135), 67-78
Abstract
Abstract The study of animal behaviour has been dominated by two general models. According to the mechanistic stimulus-response model, a particular behaviour is either an innate or an acquired habit which is simply triggered by the appropriate stimulus. By contrast, the teleological model argues that, at least, some activities are purposive actions controlled by the current value of their goals through knowledge about the instrumental relations between the actions and their consequences. The type of control over any particular behaviour can be determined by a goal revaluation procedure. If the anim al’s performance changes appropriately following an alteration in the value of the goal or...
Find it
Duan et al. (2024)
Do Membership Inference Attacks Work on Large Language Models?
In
arXiv
Abstract
Membership inference attacks (MIAs) attempt to predict whether a particular datapoint is a member of a target model's training data. Despite extensive research on traditional machine learning models, there has been limited work studying MIA on the pre-training data of large language models (LLMs). We perform a large-scale evaluation of MIAs over a suite of language models (LMs) trained on the Pile, ranging from 160M to 12B parameters. We find that MIAs barely outperform random guessing for most settings across varying LLM sizes and domains. Our further analyses reveal that this poor performance can be attributed to (1) the combination of a large dataset and few training iterations, and...
Find it
Duarte et al. (2024)
DE-COP: Detecting Copyrighted Content in Language Models Training Data
In
arXiv
Abstract
How can we detect if copyrighted content was used in the training process of a language model, considering that the training data is typically undisclosed? We are motivated by the premise that a language model is likely to identify verbatim excerpts from its training text. We propose DE-COP, a method to determine whether a piece of copyrighted content was included in training. DE-COP's core approach is to probe an LLM with multiple-choice questions, whose options include both verbatim text and their paraphrases. We construct BookTection, a benchmark with excerpts from 165 books published prior and subsequent to a model's training cutoff, along with their paraphrases. Our experiments show...
Find it
DuBrow & Davachi (2014)
Temporal Memory Is Shaped by Encoding Stability and Intervening Item Reactivation
In
The Journal of Neuroscience, 34(42), 13998-14005
Abstract
Making sense of previous experience requires remembering the order in which events unfolded in time. Prior work has implicated the hippocampus and medial temporal lobe cortex in memory for temporal information associated with individual episodes. However, the processes involved in encoding and retrieving temporal information across extended sequences is relatively poorly understood. Here we used fMRI during the encoding and retrieval of extended sequences to test specific predictions about the type of information used to resolve temporal order and the role of the hippocampus in this process. Participants studied sequences of images of celebrity faces and common objects followed by a...
Find it
Dudai (2004)
The Neurobiology of Consolidations, Or, How Stable is the Engram?
In
Annual Review of Psychology, 55(1), 51-86
Abstract
Consolidation is the progressive postacquisition stabilization of long-term memory. The term is commonly used to refer to two types of processes: synaptic consolidation, which is accomplished within the first minutes to hours after learning and occurs in all memory systems studied so far; and system consolidation, which takes much longer, and in which memories that are initially dependent upon the hippocampus undergo reorganization and may become hippocampal-independent. The textbook account of consolidation is that for any item in memory, consolidation starts and ends just once. Recently, a heated debate has been revitalized on whether this is indeed the case, or, alternatively, whether...
Find it
Duff et al. (2006)
Development of shared information in communication despite hippocampal amnesia
In
Nature Neuroscience, 9(1), 140-146
Find it
Dusek & Eichenbaum (1997)
The hippocampus and memory for orderly stimulus relations
In
Proceedings of the National Academy of Sciences, 94(13), 7109-7114
Abstract
Human declarative memory involves a systematic organization of information that supports generalizations and inferences from acquired knowledge. This kind of memory depends on the hippocampal region in humans, but the extent to which animals also have declarative memory, and whether inferential expression of memory depends on the hippocampus in animals, remains a major challenge in cognitive neuroscience. To examine these issues, we used a test of transitive inference pioneered by Piaget to assess capacities for systematic organization of knowledge and logical inference in children. In our adaptation of the test, rats were trained on a set of four overlapping odor discrimination problems...
Find it
Eakin et al. (2014)
Age invariance in semantic and episodic metamemory: Both younger and older adults provide accurate feeling-of-knowing for names of faces
In
Aging, Neuropsychology, and Cognition, 21(1), 27-51
Find it
Eich (1980)
The cue-dependent nature of state-dependent retrieval
In
Memory & Cognition, 8(2), 157-173
Abstract
An enduring problem in the study of human state-dependent retrieval concerns the apparent unpredictability of the phenomenon. Although many investigators have observed that utilization of information in episodic memory critically depends for its success on restoration, at the time of attempted retrieval, of the pharmacological state in which the information was originally acquired, many others have been unable to find evidence of such state-dependent effects. Indeed, negative results are so common that human state dependence has come to be popularly regarded as an untrustworthy phenomenon of little practical or theoretical significance. The message of this article is that the...
Find it
Eichenbaum et al. (2007)
The Medial Temporal Lobe and Recognition Memory
In
Annual Review of Neuroscience, 30(1), 123-152
Abstract
The ability to recognize a previously experienced stimulus is supported by two processes: recollection of the stimulus in the context of other information associated with the experience, and a sense of familiarity with the features of the stimulus. Although familiarity and recollection are functionally distinct, there is considerable debate about how these kinds of memory are supported by regions in the medial temporal lobes (MTL). Here, we review evidence for the distinction between recollection and familiarity and then consider the evidence regarding the neural mechanisms of these processes. Evidence from neuropsychological, neuroimaging, and neurophysiological studies of humans,...
Find it
Einstein & McDaniel (1990)
Normal aging and prospective memory.
In
Journal of Experimental Psychology: Learning, Memory, and Cognition, 16(4), 717-726
Abstract
We develop a laboratory paradigm for studying prospective memory and examine whether or not this type of memory is especially difficult for the elderly. In two experiments, young and old subjects were given a prospective memory test (they were asked to perform an action when a target event occurred) and three tests of retrospective memory (short-term memory, free recall, and recognition). From the perspective that aging disrupts mainly self-initiated retrieval processes, large age-related decrements in prospective memory were anticipated. However, despite showing reliable age differences on retrospective memory tests, both experiments showed no age deficits in prospective memory....
Find it
Einstein et al. (2005)
Multiple Processes in Prospective Memory Retrieval: Factors Determining Monitoring Versus Spontaneous Retrieval.
In
Journal of Experimental Psychology: General, 134(3), 327-342
Find it
Ellenbogen et al. (2007)
Human relational memory requires time and sleep
In
Proceedings of the National Academy of Sciences, 104(18), 7723-7728
Abstract
Relational memory, the flexible ability to generalize across existing stores of information, is a fundamental property of human cognition. Little is known, however, about how and when this inferential knowledge emerges. Here, we test the hypothesis that human relational memory develops during offline time periods. Fifty-six participants initially learned five “premise pairs” (A>B, B>C, C>D, D>E, and E>F). Unknown to subjects, the pairs contained an embedded hierarchy (A>B>C>D>E>F). Following an offline delay of either 20 min, 12 hr (wake or sleep), or 24 hr, knowledge of the hierarchy was tested by examining inferential judgments for novel “inference pairs” (B>D, C>E, and B>E). Despite...
Find it
Engle et al. (1999)
Working memory, short-term memory, and general fluid intelligence: A latent-variable approach.
In
Journal of Experimental Psychology: General, 128(3), 309-331
Find it
Ezzyat & Davachi (2014)
Similarity Breeds Proximity: Pattern Similarity within and across Contexts Is Related to Later Mnemonic Judgments of Temporal Proximity
In
Neuron, 81(5), 1179-1189
Find it
Fang et al. (2026)
Memp: Exploring Agent Procedural Memory
In
Findings of the Association for Computational Linguistics: ACL 2026, 17490-17502
Find it
Fazio et al. (2015)
Knowledge does not protect against illusory truth.
In
Journal of Experimental Psychology: General, 144(5), 993-1002
Find it
Feher da Silva & Hare (2020)
Humans primarily use model-based inference in the two-stage task
In
Nature Human Behaviour, 4(10), 1053-1066
Find it
Feng et al. (2026)
Know It, Act on It: Investigating Memory Utilization in LLM Personalization
In
arXiv
Abstract
As large language model (LLM) agents evolve into personalized companions, memory has emerged as a core capability. However, LLMs face a knowledge utilization problem: they may fail to act on relevant user preferences even when they are fully present in context. When an agent fails to tailor its response in a context where previously shared user preferences should matter, it is unclear whether the model failed to remember that information or remembered it but failed to use it. To isolate this breakdown, we introduce a decoupled evaluation paradigm that administers paired Know and Act tests to the same user preference. We conduct large-scale experiments across 16 systems and five memory...
Find it
Fleischer et al. (2012)
Allergic Reactions to Foods in Preschool-Aged Children in a Prospective Observational Food Allergy Study
In
Pediatrics, 130(1), e25-e32
Abstract
OBJECTIVE: To examine circumstances of allergic reactions to foods in a cohort of preschool-aged children. METHODS: We conducted a prospective, 5-site observational study of 512 infants aged 3 to 15 months with documented or likely allergy to milk or egg, and collected data prospectively examining allergic reactions. RESULTS: Over a median follow-up of 36 months (range: 0–48.4), the annualized reaction rate was 0.81 per year (367/512 subjects reporting 1171 reactions [95% confidence interval: 0.76–0.85]). Overall, 269/512 (52.5%) reported >1 reaction. The majority of reactions (71.2%) were triggered by milk (495 [42.3%]), egg (246 [21.0%]), and peanut (93 [7.9%]), with accidental...
Find it
Fleming et al. (2016)
Metacognition about the past and future: quantifying common and distinct influences on prospective and retrospective judgments of self-performance
In
Neuroscience of Consciousness, 2016(1), niw018
Find it
Frankland & Bontempi (2005)
The organization of recent and remote memories
In
Nature Reviews Neuroscience, 6(2), 119-130
Find it
Friedman (1993)
Memory for the time of past events.
In
Psychological Bulletin, 113(1), 44-66
Abstract
Paraphrase, not the publisher's wording. The review surveys laboratory and autobiographical work on adults' memory for when past events occurred and sets out the main established phenomena. It distinguishes several kinds of information such knowledge could rest on: distances, locations, and relative times of occurrence. Existing theories are classified in those terms and each is weighed against the available evidence. Against the common intuition that chronology is a basic property of autobiographical memory, the review concludes that human memory holds no single natural temporal code, and that a chronological past instead depends on active, repeated construction.
Find it
Fullana et al. (2016)
Neural signatures of human fear conditioning: an updated and extended meta-analysis of fMRI studies
In
Molecular Psychiatry, 21(4), 500-508
Find it
Funahashi et al. (1989)
Mnemonic coding of visual space in the monkey's dorsolateral prefrontal cortex
In
Journal of Neurophysiology, 61(2), 331-349
Abstract
1. An oculomotor delayed-response task was used to examine the spatial memory functions of neurons in primate prefrontal cortex. Monkeys were trained to fixate a central spot during a brief presentation (0.5 s) of a peripheral cue and throughout a subsequent delay period (1-6 s), and then, upon the extinction of the fixation target, to make a saccadic eye movement to where the cue had been presented. Cues were usually presented in one of eight different locations separated by 45 degrees. This task thus requires monkeys to direct their gaze to the location of a remembered visual cue, controls the retinal coordinates of the visual cues, controls the monkey's oculomotor behavior during the...
Find it
Galatzer-Levy et al. (2024)
The Cognitive Capabilities of Generative AI: A Comparative Analysis with Human Benchmarks
In
arXiv
Abstract
There is increasing interest in tracking the capabilities of general intelligence foundation models. This study benchmarks leading large language models and vision language models against human performance on the Wechsler Adult Intelligence Scale (WAIS-IV), a comprehensive, population-normed assessment of underlying human cognition and intellectual abilities, with a focus on the domains of VerbalComprehension (VCI), Working Memory (WMI), and Perceptual Reasoning (PRI). Most models demonstrated exceptional capabilities in the storage, retrieval, and manipulation of tokens such as arbitrary sequences of letters and numbers, with performance on the Working Memory Index (WMI) greater or equal...
Find it
Gamboa et al. (2022)
Obliviate! Reviewing Neural Fundamentals of Intentional Forgetting from a Meta-Analytic Perspective
In
Biomedicines, 10(7), 1555
Abstract
Intentional forgetting (IF) is an important adaptive mechanism necessary for correct memory functioning, optimal psychological wellbeing, and appropriate daily performance. Due to its complexity, the neuropsychological processes that give birth to successful intentional forgetting are not yet clearly known. In this study, we used two different meta-analytic algorithms, Activation Likelihood Estimation (ALE) & Latent Dirichlet Allocation (LDA) to quantitatively assess the neural correlates of IF and to evaluate the degree of compatibility between the proposed neurobiological models and the existing brain imaging data. We found that IF involves the interaction of two networks, the main...
Find it
Garrison et al. (2017)
Monitoring what is real: The effects of modality and action on accuracy and type of reality monitoring error
In
Cortex, 87, 108-117
Find it
Garrison et al. (2017)
Testing continuum models of psychosis: No reduction in source monitoring ability in healthy individuals prone to auditory hallucinations
In
Cortex, 91, 197-207
Find it
Garrison et al. (2015)
Paracingulate sulcus morphology is associated with hallucinations in the human brain
In
Nature Communications, 6(1)
Abstract
AbstractHallucinations are common in psychiatric disorders, and are also experienced by many individuals who are not mentally ill. Here, in 153 participants, we investigate brain structural markers that predict the occurrence of hallucinations by comparing patients with schizophrenia who have experienced hallucinations against patients who have not, matched on a number of demographic and clinical variables. Using both newly validated visual classification techniques and automated, data-driven methods, hallucinations were associated with specific brain morphology differences in the paracingulate sulcus, a fold in the medial prefrontal cortex, with a 1 cm reduction in sulcal length...
Find it
Geerts et al. (2025)
Relational reasoning and inductive bias in transformers and large language models
In
arXiv
Abstract
Transformer-based models have demonstrated remarkable reasoning abilities, but the mechanisms underlying relational reasoning remain poorly understood. We investigate how transformers perform \textit{transitive inference}, a classic relational reasoning behavior from psychology which elicits inference about indirectly related items (e.g., if $A > B$ and $B > C$, then $A > C$). We compare in-weights learning (IWL) and in-context learning (ICL) behaviors and mechanisms on these tasks, and fine profoundly different patterns of generalization. IWL models learn a linear embedding, which leads to transitive inference as well as other behavioral effects present in humans and animals. ICL models,...
Find it
Gekhman et al. (2025)
Inside-Out: Hidden Factual Knowledge in LLMs
In
arXiv
Abstract
This work presents a framework for assessing whether large language models (LLMs) encode more factual knowledge in their parameters than what they express in their outputs. While a few studies hint at this possibility, none has clearly defined or demonstrated this phenomenon. We first propose a formal definition of knowledge, quantifying it for a given question as the fraction of correct-incorrect answer pairs where the correct one is ranked higher. This gives rise to external and internal knowledge, depending on the information used to score individual answer candidates: either the model's observable token-level probabilities or its intermediate computations. Hidden knowledge arises when...
Find it
Ghosh & Gilboa (2014)
What is a memory schema? A historical perspective on current neuroscience literature
In
Neuropsychologia, 53, 104-114
Find it
Ghosh et al. (2014)
Schema Representation in Patients with Ventromedial PFC Lesions
In
The Journal of Neuroscience, 34(36), 12057-12070
Abstract
Human neuroimaging and animal studies have recently implicated the ventromedial prefrontal cortex (vmPFC) in memory schema, particularly in facilitating new encoding by existing schemas. In humans, the most conspicuous memory disorder following vmPFC damage is confabulation; strategic retrieval models suggest that aberrant schema activation or reinstatement plays a role in confabulation. This raises the possibility that beyond its role in schema-supported memory encoding, the vmPFC is also implicated in schema reinstatement itself. If that is the case, vmPFC lesions should lead to impaired schema-based operations, even on tasks that do not involve memory acquisition. To test this...
Find it
Gick & Holyoak (1983)
Schema induction and analogical transfer
In
Cognitive Psychology, 15(1), 1-38
Find it
Gilboa & Marlatte (2017)
Neurobiology of Schemas and Schema-Mediated Memory
In
Trends in Cognitive Sciences, 21(8), 618-631
Abstract
Schemas are superordinate knowledge structures that reflect abstracted commonalities across multiple experiences, exerting powerful influences over how events are perceived, interpreted, and remembered. Activated schema templates modulate early perceptual processing, as they get populated with specific informational instances (schema instantiation). Instantiated schemas, in turn, can enhance or distort mnemonic processing from the outset (at encoding), impact offline memory transformation and accelerate neocortical integration. Recent studies demonstrate distinctive neurobiological processes underlying schema-related learning. Interactions between the ventromedial prefrontal cortex...
Find it
Gillan et al. (2016)
Characterizing a psychiatric symptom dimension related to deficits in goal-directed control
In
eLife, 5
Abstract
Prominent theories suggest that compulsive behaviors, characteristic of obsessive-compulsive disorder and addiction, are driven by shared deficits in goal-directed control, which confers vulnerability for developing rigid habits. However, recent studies have shown that deficient goal-directed control accompanies several disorders, including those without an obvious compulsive element. Reasoning that this lack of clinical specificity might reflect broader issues with psychiatric diagnostic categories, we investigated whether a dimensional approach would better delineate the clinical manifestations of goal-directed deficits. Using large-scale online assessment of psychiatric symptoms and...
Find it
Girardeau et al. (2009)
Selective suppression of hippocampal ripples impairs spatial memory
In
Nature Neuroscience, 12(10), 1222-1223
Find it
Glanzman (2009)
Habituation in Aplysia: The Cheshire Cat of neurobiology
In
Neurobiology of Learning and Memory, 92(2), 147-154
Find it
Godden & Baddeley (1975)
CONTEXT‐DEPENDENT MEMORY IN TWO NATURAL ENVIRONMENTS: ON LAND AND UNDERWATER
In
British Journal of Psychology, 66(3), 325-331
Abstract
In a free recall experiment, divers learnt lists of words in two natural environments: on dry land and underwater, and recalled the words in either the environment of original learning, or in the alternative environment. Lists learnt underwater were best recalled underwater, and vice versa. A subsequent experiment shows that the disruption of moving from one environment to the other was unlikely to be responsible for context-dependent memory.
Find it
Gong et al. (2024)
Working Memory Capacity of ChatGPT: An Empirical Study
In
Proceedings of the AAAI Conference on Artificial Intelligence, 38(9), 10048-10056
Abstract
Working memory is a critical aspect of both human intelligence and artificial intelligence, serving as a workspace for the temporary storage and manipulation of information. In this paper, we systematically assess the working memory capacity of ChatGPT, a large language model developed by OpenAI, by examining its performance in verbal and spatial n-back tasks under various conditions. Our experiments reveal that ChatGPT has a working memory capacity limit strikingly similar to that of humans. Furthermore, we investigate the impact of different instruction strategies on ChatGPT's performance and observe that the fundamental patterns of a capacity limit persist. From our empirical findings,...
Find it
Goodwin et al. (1969)
Alcohol and Recall: State-Dependent Effects in Man
In
Science, 163(3873), 1358-1360
Abstract
Male volunteers performed four memory tasks either while sober or under effects of alcohol. Twenty-four hours later they were tested under the same or different conditions. In tasks measuring recall and interference, learning transfer was better when the subject was intoxicated during both sessions than when he was intoxicated only during the learning session. In a task measuring recognition, transfer was not significantly affected by changing state. Thus, alcohol appears to produce "dissociated" or state-dependent effects in man, but not all forms of memory are equally sensitive to the phenomenon.
Find it
Gorman et al. (2013)
What's learned together stays together: Speakers' choice of referring expression reflects shared experience.
In
Journal of Experimental Psychology: Learning, Memory, and Cognition, 39(3), 843-853
Find it
Graf et al. (1984)
The information that amnesic patients do not forget.
In
Journal of Experimental Psychology: Learning, Memory, and Cognition, 10(1), 164-178
Find it
Groves & Thompson (1970)
Habituation: A dual-process theory.
In
Psychological Review, 77(5), 419-450
Abstract
Presented a dual-process theory of response plasticity to repeated stimulation. 2 hypothetical processes, 1 decremental (habituation) and 1 incremental (sensitization), are assumed to develop independently in the CNS and interact to yield the final behavioral outcome. Behavioral experiments are presented, using both the hindlimb flexion reflex of acute spinal cat and the acoustic startle response of intact rat, which are consistent with this theory. Neurophysiological experiments indicate that the 2 processes have separate and distinct neuronal substrates. The dual-process theory and other current theories of response habituation are evaluated in terms of these and other recent findings....
Find it
Guo et al. (2026)
The Interference Gap: Comparing Retrieval Bounds in Human Memory and RAG Systems
In
arXiv
Abstract
How do retrieval bounds compare between human episodic memory and Retrieval-Augmented Generation (RAG) systems under semantic interference? We present a unified signal detection theory (SDT) framework that applies to both, and use it to fit behavioral and computational data in matched paradigms. Both systems show logarithmic accuracy decline with association count (fan), but humans exhibit lower interference sensitivity ($α/σ= 0.41$) than dense passage retrieval ($α/σ= 0.67$), with cognitively-inspired HippoRAG falling between the two ($α/σ= 0.44$). Behavioral experiments ($N = 112$) and simulations validate the framework; parameter recovery confirms identifiability ($r \geq .93$) and...
Find it
Gurnee & Tegmark (2024)
Language Models Represent Space and Time
In
arXiv
Abstract
The capabilities of large language models (LLMs) have sparked debate over whether such systems just learn an enormous collection of superficial statistics or a set of more coherent and grounded representations that reflect the real world. We find evidence for the latter by analyzing the learned representations of three spatial datasets (world, US, NYC places) and three temporal datasets (historical figures, artworks, news headlines) in the Llama-2 family of models. We discover that LLMs learn linear representations of space and time across multiple scales. These representations are robust to prompting variations and unified across different entity types (e.g. cities and landmarks). In...
Find it
Hall et al. (2014)
The Neural Basis of Involuntary Episodic Memories
In
Journal of Cognitive Neuroscience, 26(10), 2385-2399
Abstract
Voluntary episodic memories require an intentional memory search, whereas involuntary episodic memories come to mind spontaneously without conscious effort. Cognitive neuroscience has largely focused on voluntary memory, leaving the neural mechanisms of involuntary memory largely unknown. We hypothesized that, because the main difference between voluntary and involuntary memory is the controlled retrieval processes required by the former, there would be greater frontal activity for voluntary than involuntary memories. Conversely, we predicted that other components of the episodic retrieval network would be similarly engaged in the two types of memory. During encoding, all participants...
Find it
Hall et al. (2018)
Neural responses to emotional involuntary memories in posttraumatic stress disorder: Differences in timing and activity
In
NeuroImage: Clinical, 19, 793-804
Find it
Hallford et al. (2022)
Reduced specificity and increased overgenerality of autobiographical memory persist as cognitive vulnerabilities in remitted major depression: A meta‐analysis
In
Clinical Psychology & Psychotherapy, 29(5), 1515-1529
Abstract
AbstractDifficulty in accessing specific memories, referred to as reduced memory specificity or overgeneral memory (OGM), has been established as a marker of clinical depression. However, it is not clear if this deficit persists following the remission of depressive episodes. The current study involved a systematic review and meta‐analysis of empirical studies with the aim of establishing whether remitted depression was associated with retrieving fewer specific and more overgeneral autobiographical memories. Seventeen studies were identified as eligible. The results indicated that people with remitted depression recalled fewer specific memories (k = 15; g = −0.314, 95% CI [−0.543;...
Find it
Hamilton & Martin (2007)
Proactive Interference in a Semantic Short-Term Memory Deficit: Role of Semantic and Phonological Relatedness
In
Cortex, 43(1), 112-123
Find it
Hardt et al. (2013)
Decay happens: the role of active forgetting in memory
In
Trends in Cognitive Sciences, 17(3), 111-120
Find it
Hardwick et al. (2013)
A quantitative meta-analysis and review of motor learning in the human brain
In
NeuroImage, 67, 283-297
Find it
Hart (1965)
Memory and the feeling-of-knowing experience.
In
Journal of Educational Psychology, 56(4), 208-216
Find it
Hassabis et al. (2014)
Imagine All the People: How the Brain Creates and Uses Personality Models to Predict Behavior
In
Cerebral Cortex, 24(8), 1979-1987
Find it
Hawkins et al. (2020)
Generalizing meanings from partners to populations: Hierarchical inference supports convention formation on networks
In
arXiv
Abstract
A key property of linguistic conventions is that they hold over an entire community of speakers, allowing us to communicate efficiently even with people we have never met before. At the same time, much of our language use is partner-specific: we know that words may be understood differently by different people based on our shared history. This poses a challenge for accounts of convention formation. Exactly how do agents make the inferential leap to community-wide expectations while maintaining partner-specific knowledge? We propose a hierarchical Bayesian model to explain how speakers and listeners solve this inductive problem. To evaluate our model's predictions, we conducted an...
Find it
Hawkins et al. (2023)
From partners to populations: A hierarchical Bayesian account of coordination and convention.
In
Psychological Review, 130(4), 977-1016
Find it
Healey et al. (2019)
Contiguity in episodic memory
In
Psychonomic Bulletin & Review, 26(3), 699-720
Find it
Hebscher et al. (2016)
Memory, Decision-Making, and the Ventromedial Prefrontal Cortex (vmPFC): The Roles of Subcallosal and Posterior Orbitofrontal Cortices in Monitoring and Control Processes
In
Cerebral Cortex, 26(12), 4590-4601
Find it
Henry et al. (2004)
A Meta-Analytic Review of Prospective Memory and Aging.
In
Psychology and Aging, 19(1), 27-39
Abstract
A meta-analysis of prospective memory (PM) studies revealed that in laboratory settings younger participants outperform older participants on tests of both time- and event-based PM (rs=-.39 and -.34, respectively). Event-based PM tasks that impose higher levels of controlled strategic demand are associated with significantly larger age effects than event-based PM tasks that are supported by relatively more automatic processes (rs=-.40 vs. -.14, respectively). However, contrary to the prevailing view in the literature, retrospective memory as measured by free recall is associated with significantly greater age-related decline (r=-.52) than PM, and older participants perform substantially...
Find it
Higham & Higham (2019)
New improved gamma: Enhancing the accuracy of Goodman–Kruskal’s gamma using ROC curves
In
Behavior Research Methods, 51(1), 108-125
Find it
Holdstock et al. (2002)
Under what conditions is recognition spared relative to recall after selective hippocampal damage in humans?
In
Hippocampus, 12(3), 341-351
Abstract
Abstract The claim that recognition memory is spared relative to recall after focal hippocampal damage has been disputed in the literature. We examined this claim by investigating object and object–location recall and recognition memory in a patient, YR, who has adult‐onset selective hippocampal damage. Our aim was to identify the conditions under which recognition was spared relative to recall in this patient. She showed unimpaired forced‐choice object recognition but clearly impaired recall, even when her control subjects found the object recognition task to be numerically harder than the object recall task. However, on two other recognition tests, YR's performance was not relatively...
Find it
Horton & Gerrig (2005)
The impact of memory demands on audience design during language production
In
Cognition, 96(2), 127-142
Abstract
Speakers often tailor their utterances to the needs of particular addressees—a process called audience design. We argue that important aspects of audience design can be understood as emergent features of ordinary memory processes. This perspective contrasts with earlier views that presume special processes or representations. To support our account, we present a study in which Directors engaged in a referential communication task with two independent Matchers. Over several rounds, the Directors instructed the Matchers how to arrange a set of picture cards. For half the triads, the Directors' card categories were initially distributed orthogonally by Matcher (e.g. Directors described birds...
Find it
Hu et al. (2020)
Promoting memory consolidation during sleep: A meta-analysis of targeted memory reactivation.
In
Psychological Bulletin, 146(3), 218-244
Abstract
Targeted memory reactivation (TMR) is a methodology employed to manipulate memory processing during sleep. TMR studies have great potential to advance understanding of sleep-based memory consolidation and corresponding neural mechanisms. Research making use of TMR has developed rapidly, with over 70 articles published in the last decade, yet no quantitative analysis exists to evaluate the overall effects. Here we present the first meta-analysis of sleep TMR, compiled from 91 experiments with 212 effect sizes (N= 2,004). Based on multilevel modeling, overall sleep TMR was highly effective (Hedges’g= 0.29, 95% CI \[0.21, 0.38\]), with a significant effect for two stages of...
Find it
Hu & Lewis (2025)
Do Language Models Understand the Cognitive Tasks Given to Them? Investigations with the N-Back Paradigm
In
arXiv
Abstract
Cognitive tasks originally developed for humans are now increasingly used to study language models. While applying these tasks is often straightforward, interpreting their results can be challenging. In particular, when a model underperforms, it is often unclear whether this results from a limitation in the cognitive ability being tested or a failure to understand the task itself. A recent study argues that GPT 3.5's declining performance on 2-back and 3-back tasks reflects a working memory capacity limit similar to humans (Gong et al., 2024). By analyzing a range of open-source language models of varying performance levels on these tasks, we show that the poor performance is due at least...
Find it
Hu et al. (2025)
Evaluating Memory in LLM Agents via Incremental Multi-Turn Interactions
In
arXiv
Abstract
Recent benchmarks for Large Language Model (LLM) agents primarily focus on evaluating reasoning, planning, and execution capabilities, while another critical component-memory, encompassing how agents memorize, update, and retrieve long-term information-is under-evaluated due to the lack of benchmarks. We term agents with memory mechanisms as memory agents. In this paper, based on classic theories from memory science and cognitive science, we identify four core competencies essential for memory agents: accurate retrieval, test-time learning, long-range understanding, and selective forgetting. Existing benchmarks either rely on limited context lengths or are tailored for static,...
Find it
Hua & Artzi (2024)
Talk Less, Interact Better: Evaluating In-context Conversational Adaptation in Multimodal LLMs
In
arXiv
Abstract
Humans spontaneously use increasingly efficient language as interactions progress, by adapting and forming ad-hoc conventions. This phenomenon has been studied extensively using reference games, showing properties of human language that go beyond relaying intents. It remains unexplored whether multimodal large language models (MLLMs) similarly increase communication efficiency during interactions, and what mechanisms they may adopt for this purpose. We introduce ICCA, an automated framework to evaluate such conversational adaptation as an in-context behavior in MLLMs. We evaluate several state-of-the-art MLLMs, and observe that while they may understand the increasingly efficient language...
Find it
Huang et al. (2025)
On the Failure of Latent State Persistence in Large Language Models
In
arXiv
Abstract
While Large Language Models (LLMs) exhibit remarkable reasoning abilities, we demonstrate that they fundamentally lack a core aspect of human cognition: working memory. Human working memory is an active cognitive system that enables not only the temporary storage of information but also its processing and utilization. Without working memory, individuals may produce unrealistic conversations, exhibit self-contradictions, and struggle with tasks that require mental reasoning. Existing evaluations using N-back or context-dependent tasks fail as they allow LLMs to exploit external context rather than retaining the reasoning process in the latent space. We introduce three novel tasks—(1)...
Find it
Huet et al. (2025)
Episodic Memories Generation and Evaluation Benchmark for Large Language Models
In
arXiv
Abstract
Episodic memory -- the ability to recall specific events grounded in time and space -- is a cornerstone of human cognition, enabling not only coherent storytelling, but also planning and decision-making. Despite their remarkable capabilities, Large Language Models (LLMs) lack a robust mechanism for episodic memory: we argue that integrating episodic memory capabilities into LLM is essential for advancing AI towards human-like cognition, increasing their potential to reason consistently and ground their output in real-world episodic events, hence avoiding confabulations. To address this challenge, we introduce a comprehensive framework to model and evaluate LLM episodic memory...
Find it
Hutchison et al. (2013)
The semantic priming project
In
Behavior Research Methods, 45(4), 1099-1114
Find it
Jaeggi et al. (2010)
The concurrent validity of the N-back task as a working memory measure
In
Memory, 18(4), 394-412
Abstract
The N-back task is used extensively in literature as a working memory (WM) paradigm and it is increasingly used as a measure of individual differences. However, not much is known about the psychometric properties of this task and the current study aims to shed more light on this issue. We first review the current literature on the psychometric properties of the N-back task. With three experiments using task variants with different stimuli and load levels, we then investigate the nature of the N-back task by investigating its relationship to WM, and its role as an inter-individual difference measure. Consistent with previous literature, our data suggest that the N-back task is not a useful...
Find it
Jagielski et al. (2023)
Measuring Forgetting of Memorized Training Examples
In
arXiv
Abstract
Machine learning models exhibit two seemingly contradictory phenomena: training data memorization, and various forms of forgetting. In memorization, models overfit specific training examples and become susceptible to privacy attacks. In forgetting, examples which appeared early in training are forgotten by the end. In this work, we connect these phenomena. We propose a technique to measure to what extent models "forget" the specifics of training examples, becoming less susceptible to privacy attacks on examples they have not seen recently. We show that, while non-convex models can memorize data forever in the worst-case, standard image, speech, and language models empirically do forget...
Find it
Janacsek et al. (2020)
Sequence learning in the human brain: A functional neuroanatomical meta-analysis of serial reaction time studies
In
NeuroImage, 207, 116387
Find it
Janik (2023)
Aspects of human memory and Large Language Models
In
arXiv
Abstract
Large Language Models (LLMs) are huge artificial neural networks which primarily serve to generate text, but also provide a very sophisticated probabilistic model of language use. Since generating a semantically consistent text requires a form of effective memory, we investigate the memory properties of LLMs and find surprising similarities with key characteristics of human memory. We argue that the human-like memory properties of the Large Language Model do not follow automatically from the LLM architecture but are rather learned from the statistics of the training textual data. These results strongly suggest that the biological features of human memory leave an imprint on the way that...
Find it
Janowsky et al. (1989)
Source memory impairment in patients with frontal lobe lesions
In
Neuropsychologia, 27(8), 1043-1056
Find it
Janssen et al. (2006)
Memory for time: How people date events
In
Memory & Cognition, 34(1), 138-147
Find it
Ji-An et al. (2024)
Linking In-context Learning in Transformers to Human Episodic Memory
In
arXiv
Abstract
Understanding connections between artificial and biological intelligent systems can reveal fundamental principles of general intelligence. While many artificial intelligence models have a neuroscience counterpart, such connections are largely missing in Transformer models and the self-attention mechanism. Here, we examine the relationship between interacting attention heads and human episodic memory. We focus on induction heads, which contribute to in-context learning in Transformer-based large language models (LLMs). We demonstrate that induction heads are behaviorally, functionally, and mechanistically similar to the contextual maintenance and retrieval (CMR) model of human episodic...
Find it
Jin et al. (2022)
Across-subject correlation between confidence and accuracy: A meta-analysis of the Confidence Database
In
Psychonomic Bulletin & Review, 29(4), 1405-1413
Find it
Johnson & Raye (1981)
Reality monitoring.
In
Psychological Review, 88(1), 67-85
Abstract
People remember information from 2 basic sources: external (perceptual processes) and internal (processes such as reasoning, imagination, and thought). Of particular interest are the processes people use in deciding whether information initially had an external or an internal source, i.e., "reality monitoring." A working model of reality monitoring is proposed to account for both discrimination and confusion between memories for thoughts and memories for perceptions. Examples of questions the model addresses are: What types of information are more likely to be represented in memories of external events than in memories of internal events?; What assumptions do individuals have about their...
Find it
Johnson et al. (1988)
Phenomenal characteristics of memories for perceived and imagined autobiographical events.
In
Journal of Experimental Psychology: General, 117(4), 371-376
Find it
Johnson et al. (1993)
Source monitoring.
In
Psychological Bulletin, 114(1), 3-28
Abstract
A framework for understanding source monitoring and relevant empirical evidence is described, and several related phenomena are discussed: old–new recognition, indirect tests, eyewitness testimony, misattributed familiarity, cryptomnesia, and incorporation of fiction into fact. Disruptions in source monitoring (e.g., from confabulation, amnesia, and aging) and the brain regions that are involved are also considered, and source monitoring within a general memory architecture is discussed. It is argued that source monitoring is based on qualities of experience resulting from combinations of perceptual and reflective processes, usually requires relatively differentiated phenomenal...
Find it
Jones et al. (2026)
LLMs and people both learn to form conventions -- just not with each other
In
arXiv
Abstract
Humans align to one another in conversation -- adopting shared conventions that ease communication. We test whether LLMs form the same kinds of conventions in a multimodal communication game. Both humans and LLMs display evidence of convention-formation (increasing the accuracy and consistency of their turns while decreasing their length) when communicating in same-type dyads (humans with humans, AI with AI). However, heterogenous human-AI pairs fail -- suggesting differences in communicative tendencies. In Experiment 2, we ask whether LLMs can be induced to behave more like human conversants, by prompting them to produce superficially humanlike behavior. While the length of their...
Find it
Jonides & Nee (2006)
Brain mechanisms of proactive interference in working memory
In
Neuroscience, 139(1), 181-193
Find it
Jumelet et al. (2024)
Do Language Models Exhibit Human-like Structural Priming Effects?
In
Findings of the Association for Computational Linguistics ACL 2024, 14727-14742
Abstract
We explore which linguistic factors—at the sentence and token level—play an important role in influencing language model predictions, and investigate whether these are reflective of results found in humans and human corpora (Gries and Kootstra, 2017). We make use of the structural priming paradigm—where recent exposure to a structure facilitates processing of the same structure—to investigate where priming effects manifest, and what factors predict them. We find these effects can be explained via the inverse frequency effect found in human priming, where rarer elements within a prime increase priming effects, as well as lexical dependence between prime and target. Our results provide an...
Find it
Juslin et al. (2000)
Naive empiricism and dogmatism in confidence research: A critical examination of the hard–easy effect.
In
Psychological Review, 107(2), 384-396
Abstract
Two robust phenomena in research on confidence in one's general knowledge are the overconfidence phenomenon and the hard-easy effect. In this article, the authors propose that the hard-easy effect has been interpreted with insufficient attention to the scale-end effects, the linear dependency, and the regression effects in data and that the continued adherence to the idea of a "cognitive overconfidence bias" is mediated by selective attention to particular data sets. A quantitative review of studies with 2-alternative general knowledge items demonstrates that, contrary to widespread belief, there is (a) very little support for a cognitive-processing bias in these data; (b) a difference...
Find it
Kadavath et al. (2022)
Language Models (Mostly) Know What They Know
In
arXiv
Abstract
We study whether language models can evaluate the validity of their own claims and predict which questions they will be able to answer correctly. We first show that larger models are well-calibrated on diverse multiple choice and true/false questions when they are provided in the right format. Thus we can approach self-evaluation on open-ended sampling tasks by asking models to first propose answers, and then to evaluate the probability "P(True)" that their answers are correct. We find encouraging performance, calibration, and scaling for P(True) on a diverse array of tasks. Performance at self-evaluation further improves when we allow models to consider many of their own samples before...
Find it
Kail (2002)
Developmental Change in Proactive Interference
In
Child Development, 73(6), 1703-1714
Abstract
Abstract The aim of the present research was to examine age–related change in proactive interference, which refers to impaired recall due to interference from material presented previously. Study 1 was a meta–analysis based on 26 studies that included 82 data sets. The results indicated that proactive interference decreased between 4 and 13 years of age. In Study 2, children from grades 3 through 6 and college students (N= 125) were administered a short–term memory task in which they briefly remembered sets of three words. For all ages, recall was accurate on Trial 1. However, recall became less accurate over Trials 2 through 4, particularly for the younger children in the sample. In...
Find it
Kale (2026)
Future Confidence Distillation in Large Language Models
In
arXiv
Abstract
Reliable confidence estimation is essential for deploying large language models (LLMs) in confidence-aware systems, where downstream decisions such as retrieval, tool use, and adaptive computation depend on accurately estimating answer reliability. Existing approaches, however, largely treat confidence as a property of completed responses, overlooking how confidence-related information evolves throughout the answering process. In this work, we investigate confidence from a temporal perspective by comparing pre-solution Feeling-of-Knowing (FOK) and post-solution Judgement-of-Learning (JOL) confidence estimates across frontier and open-source LLMs. We show that post-solution confidence is...
Find it
Kalra et al. (2019)
Evidence of stable individual differences in implicit learning
In
Cognition, 190, 199-211
Find it
Kandpal et al. (2023)
Large Language Models Struggle to Learn Long-Tail Knowledge
In
arXiv
Abstract
The Internet contains a wealth of knowledge -- from the birthdays of historical figures to tutorials on how to code -- all of which may be learned by language models. However, while certain pieces of information are ubiquitous on the web, others appear extremely rarely. In this paper, we study the relationship between the knowledge memorized by large language models and the information in pre-training datasets scraped from the web. In particular, we show that a language model's ability to answer a fact-based question relates to how many documents associated with that question were seen during pre-training. We identify these relevant documents by entity linking pre-training datasets and...
Find it
Kang et al. (2025)
In-Context Learning can Perform Continual Learning Like Humans
In
arXiv
Abstract
Large language models (LLMs) can adapt to new tasks via in-context learning (ICL) without parameter updates, making them powerful learning engines for fast adaptation. While extensive research has examined ICL as a few-shot learner, whether it can achieve long-term retention and cross-task knowledge accumulation when multitasks arrive sequentially remains underexplored. Motivated by human memory studies, we investigate the retention characteristics of ICL in multitask settings and extend it to in-context continual learning (ICCL), where continual learning ability emerges through task scheduling and prompt rearrangement. Experiments on Markov-Chain benchmarks demonstrate that, for specific...
Find it
Kauf et al. (2023)
Event Knowledge in Large Language Models: The Gap Between the Impossible and the Unlikely
In
Cognitive Science, 47(11)
Abstract
AbstractWord co‐occurrence patterns in language corpora contain a surprising amount of conceptual knowledge. Large language models (LLMs), trained to predict words in context, leverage these patterns to achieve impressive performance on diverse semantic tasks requiring world knowledge. An important but understudied question about LLMs’ semantic abilities is whether they acquire generalized knowledge of common events. Here, we test whether five pretrained LLMs (from 2018's BERT to 2023's MPT) assign a higher likelihood to plausible descriptions of agent−patient interactions than to minimally different implausible versions of the same event. Using three curated sets of minimal sentence...
Find it
Kausche et al. (2025)
Fear and safety learning in anxiety- and stress-related disorders: An updated meta-analysis
In
Neuroscience & Biobehavioral Reviews, 169, 105983
Find it
Khalifa et al. (2024)
Source-Aware Training Enables Knowledge Attribution in Language Models
In
arXiv
Abstract
Large language models (LLMs) learn a vast amount of knowledge during pretraining, but they are often oblivious to the source(s) of such knowledge. We investigate the problem of intrinsic source citation, where LLMs are required to cite the pretraining source supporting a generated response. Intrinsic source citation can enhance LLM transparency, interpretability, and verifiability. To give LLMs such ability, we explore source-aware training -- a recipe that involves (i) training the LLM to associate unique source document identifiers with the knowledge in each document, followed by (ii) an instruction-tuning stage to teach the LLM to cite a supporting pretraining source when prompted....
Find it
Kikyo et al. (2002)
Neural Correlates for Feeling-of-Knowing
In
Neuron, 36(1), 177-186
Find it
Kim (2011)
Neural activity that predicts subsequent memory and forgetting: A meta-analysis of 74 fMRI studies
In
NeuroImage, 54(3), 2446-2461
Find it
Kindt et al. (2009)
Beyond extinction: erasing human fear responses and preventing the return of fear
In
Nature Neuroscience, 12(3), 256-258
Abstract
Animal studies have shown that fear memories can change when recalled, a process referred to as reconsolidation. We found that oral administration of the β-adrenergic receptor antagonist propranolol before memory reactivation in humans erased the behavioral expression of the fear memory 24 h later and prevented the return of fear. Disrupting the reconsolidation of fear memory opens up new avenues for providing a long-term cure for patients with emotional disorders.
Find it
Kirichenko et al. (2025)
AbstentionBench: Reasoning LLMs Fail on Unanswerable Questions
In
arXiv
Abstract
For Large Language Models (LLMs) to be reliably deployed in both everyday and high-stakes domains, knowing when not to answer is equally critical as answering correctly. Real-world user queries, which can be underspecified, ill-posed, or fundamentally unanswerable, require LLMs to reason about uncertainty and selectively abstain -- i.e., refuse to answer definitively. However, abstention remains understudied, without a systematic evaluation framework for modern LLMs. In this work, we introduce AbstentionBench, a large-scale benchmark for holistically evaluating abstention across 20 diverse datasets, including questions with unknown answers, underspecification, false premises, subjective...
Find it
Kliegel et al. (2008)
Adult age differences in event-based prospective memory: A meta-analysis on the role of focal versus nonfocal cues.
In
Psychology and Aging, 23(1), 203-208
Find it
Klingelhöfer-Jens et al. (2022)
Robust group- but limited individual-level (longitudinal) reliability and insights into cross-phases response prediction of conditioned fear
In
eLife, 11
Abstract
Here, we follow the call to target measurement reliability as a key prerequisite for individual-level predictions in translational neuroscience by investigating (1) longitudinal reliability at the individual and (2) group level, (3) internal consistency and (4) response predictability across experimental phases. One hundred and twenty individuals performed a fear conditioning paradigm twice 6 months apart. Analyses of skin conductance responses, fear ratings and blood oxygen level dependent functional magnetic resonance imaging (BOLD fMRI) with different data transformations and included numbers of trials were conducted. While longitudinal reliability was rather limited at the individual...
Find it
Knowlton et al. (1996)
A Neostriatal Habit Learning System in Humans
In
Science, 273(5280), 1399-1402
Abstract
Amnesic patients and nondemented patients with Parkinson's disease were given a probabilistic classification task in which they learned which of two outcomes would occur on each trial, given the particular combination of cues that appeared. Amnesic patients exhibited normal learning of the task but had severely impaired declarative memory for the training episode. In contrast, patients with Parkinson's disease failed to learn the probabilistic classification task, despite having intact memory for the training episode. This double dissociation shows that the limbic-diencephalic regions damaged in amnesia and the neostriatum damaged in Parkinson's disease support separate and parallel...
Find it
Koen & Yonelinas (2014)
The Effects of Healthy Aging, Amnestic Mild Cognitive Impairment, and Alzheimer’s Disease on Recollection and Familiarity: A Meta-Analytic Review
In
Neuropsychology Review, 24(3), 332-354
Find it
Kompus et al. (2011)
Multimodal Imaging of Incidental Retrieval: The Low Route to Memory
In
Journal of Cognitive Neuroscience, 23(4), 947-960
Abstract
AbstractMemories of past episodes frequently come to mind incidentally, without directed search. It has remained unclear how incidental retrieval processes are initiated in the brain. Here we used fMRI and ERP recordings to find brain activity that specifically correlates with incidental retrieval, as compared to intentional retrieval. Intentional retrieval was associated with increased activation in dorsolateral prefrontal cortex. By contrast, incidental retrieval was associated with a reduced fMRI signal in posterior brain regions, including extrastriate and parahippocampal cortex, and a modulation of a posterior ERP component 170 msec after the onset of visual retrieval cues....
Find it
Koriat (1993)
How do we know that we know? The accessibility model of the feeling of knowing.
In
Psychological Review, 100(4), 609-639
Abstract
Even when Ss fail to recall a solicited target, they can provide feeling-of-knowing (FOK) judgments about its availability in memory. Most previous studies addressed the question of FOK accuracy, only a few examined how FOK itself is determined, and none asked how the processes assumed to underlie FOK also account for its accuracy. The present work examined all 3 questions within a unified model, with the aim of demystifying the FOK phenomenon. The model postulates that the computation of FOK is parasitic on the processes involved in attempting to retrieve the target, relying on the accessibility of pertinent information. It specifies the links between memory strength, accessibility of...
Find it
Koriat & Goldsmith (1996)
Monitoring and control processes in the strategic regulation of memory accuracy.
In
Psychological Review, 103(3), 490-517
Abstract
When people are allowed freedom to volunteer or withhold information, they can enhance the accuracy of their memory reports substantially relative to forced-report performance. A theoretical framework addressing the strategic regulation of memory reporting is put forward that delineates the mediating role of metamemorial monitoring and control processes. Although the enhancement of memory accuracy is generally accompanied by a reduction in memory quantity, experimental and simulation results indicate that both of these effects depend critically on (a) accuracy incentive and ( b ) monitoring effectiveness. The results are discussed with regard to the contribution of metamemory processes to...
Find it
Koriat (2012)
The self-consistency model of subjective confidence.
In
Psychological Review, 119(1), 80-113
Abstract
How do people monitor the correctness of their answers? A self-consistency model is proposed for the process underlying confidence judgments and their accuracy. In answering a 2-alternative question, participants are assumed to retrieve a sample of representations of the question and base their confidence on the consistency with which the chosen answer is supported across representations. Confidence is modeled by analogy to the calculation of statistical level of confidence (SLC) in testing hypotheses about a population and represents the participant's assessment of the likelihood that a new sample will yield the same choice. Assuming that participants draw representations from a commonly...
Find it
Kredlow et al. (2016)
Harnessing reconsolidation to weaken fear and appetitive memories: A meta-analysis of post-retrieval extinction effects.
In
Psychological Bulletin, 142(3), 314-336
Abstract
A new understanding of the mechanisms of memory retrieval and reconsolidation holds the potential for improving exposure-based treatments. Basic research indicates that following fear extinction, safety and fear memories may compete, raising the possibility of return of fear. One possible solution is to modify original fear memories through reconsolidation interference, reducing the likelihood of return of fear. Postretrieval extinction is a behavioral method of reconsolidation interference that has been explored in the context of conditioned fear and appetitive memory paradigms. This meta-analysis examines the magnitude of postretrieval extinction effects and potential moderators of...
Find it
Krishnamurthy et al. (2024)
Can large language models explore in-context?
In
arXiv
Abstract
We investigate the extent to which contemporary Large Language Models (LLMs) can engage in exploration, a core capability in reinforcement learning and decision making. We focus on native performance of existing LLMs, without training interventions. We deploy LLMs as agents in simple multi-armed bandit environments, specifying the environment description and interaction history entirely in-context, i.e., within the LLM prompt. We experiment with GPT-3.5, GPT-4, and Llama2, using a variety of prompt designs, and find that the models do not robustly engage in exploration without substantial interventions: i) Across all of our experiments, only one configuration resulted in satisfactory...
Find it
Kronmüller & Barr (2015)
Referential precedents in spoken language comprehension: A review and meta-analysis
In
Journal of Memory and Language, 83, 1-19
Find it
Kwok et al. (2025)
Neural Representation of Episodic Time
In
The Journal of Neuroscience, 45(46), e1397252025
Abstract
Inspired by recent discoveries of neural populations that track time for specific moments (time cells) and elapsed durations (temporal context and periodic time cells), this review, based on a minisymposium presented at the 2025 annual meeting of the Society for Neuroscience, brings together macro-, meso-, and micro-scale neural evidence to formulate hypotheses about how episodic time—the tracking and organizing of events in time as we experience, store, and retrieve them from memory—is coded and processed in the mammalian brain. We also discuss computational principles and relationships to other related phenomena such as memory replay and emotional states.
Find it
Kwon et al. (2023)
A meta-analysis of event-related potential correlates of recognition memory
In
Psychonomic Bulletin & Review, 30(6), 2083-2105
Abstract
Abstract A longstanding question in memory research is whether recognition is supported by more than one mnemonic process. Dual-process models distinguish recollection of episodic detail from familiarity, while single-process models explain recognition in terms of one process that varies in strength. Dual process models have drawn support from findings that recollection and familiarity elicit distinct electroencephalographic event-related potentials (ERPs): a mid-frontal ERP effect that occurs at around 300–500 ms post-stimulus onset and is often larger for familiarity than recollection contrasts, and a parietal ERP effect that occurs at around 500–800 ms and is larger for recollection...
Find it
La Voie & Light (1994)
Adult age differences in repetition priming: A meta-analysis.
In
Psychology and Aging, 9(4), 539-553
Find it
Laing et al. (2026)
Emotional learning selectively distorts the temporal organization of memory: A quantitative synthesis
In
Cognition, 268, 106382
Find it
Ralph et al. (2017)
The neural and computational bases of semantic cognition
In
Nature Reviews Neuroscience, 18(1), 42-55
Abstract
Semantic cognition refers to our ability to use, manipulate and generalize knowledge that is acquired over the lifespan to support innumerable verbal and non-verbal behaviours. This Review summarizes key findings and issues arising from a decade of research into the neurocognitive and neurocomputational underpinnings of this ability, leading to a new framework that we term controlled semantic cognition (CSC). CSC offers solutions to long-standing queries in philosophy and cognitive science, and yields a convergent framework for understanding the neural and computational bases of healthy semantic cognition and its dysfunction in brain disorders.
Find it
Lampinen et al. (2001)
Recollections of things schematic: Room schemas revisited.
In
Journal of Experimental Psychology: Learning, Memory, and Cognition, 27(5), 1211-1222
Find it
Lavallé et al. (2023)
The neural signature of reality‐monitoring: A meta‐analysis of functional neuroimaging studies
In
Human Brain Mapping, 44(11), 4372-4389
Abstract
Abstract Distinguishing imagination and thoughts from information we perceived from the environment, a process called reality‐monitoring, is important in everyday situations. Although reality monitoring seems to overlap with the concept of self‐monitoring, which allows one to distinguish self‐generated actions or thoughts from those generated by others, the two concepts remain largely separate cognitive domains and their common brain substrates have received little attention. We investigated the brain regions involved in these two cognitive processes and explored the common brain regions they share. To do this, we conducted two separate coordinate‐based meta‐analyses of functional...
Find it
Lee et al. (2017)
An Update on Memory Reconsolidation Updating
In
Trends in Cognitive Sciences, 21(7), 531-545
Find it
Lee et al. (2026)
Do Language Models Show Structural Priming Across Different Domains?
In
Proceedings of the 1st Workshop on Computational Developmental Linguistics (CDL), 37-51
Find it
Lefebvre et al. (2017)
Behavioural and neural characterization of optimistic reinforcement learning
In
Nature Human Behaviour, 1(4)
Find it
Levine et al. (2002)
Aging and autobiographical memory: Dissociating episodic from semantic retrieval.
In
Psychology and Aging, 17(4), 677-689
Abstract
Cognitive aging research documents reduced access to contextually specific episodic details in older adults, whereas access to semantic or other nonepisodic information is preserved or facilitated. The present study extended this finding to autobiographical memory by using a new measure: the Autobiographical Interview. Younger and older adults recalled events from 5 life periods. Protocols were scored according to a reliable system for categorizing episodic and nonepisodic information. Whereas younger adults were biased toward episodic details reflecting happenings, locations, perceptions, and thoughts, older adults favored semantic details not connected to a particular time and place....
Find it
Lindsay & Johnson (1989)
The eyewitness suggestibility effect and memory for source
In
Memory & Cognition, 17(3), 349-358
Find it
Lindsey (2025)
Emergent Introspective Awareness in Large Language Models
In
Transformer Circuits Thread
Find it
Link et al. (2026)
The Connection Between Associative Memory and Semantic Similarity: Evidence From Fan Experiments and Distributional Models
In
Cognitive Science, 50(4)
Abstract
Abstract Memory retrieval is prone to interference: when multiple concepts in memory match a given retrieval cue, recall becomes slower and less accurate. This has repeatedly been studied in fan effect experiments in which participants learn facts that are combinations of person–location pairs. These experiments manipulate the fan of a concept—the number of facts linked to it—establishing interference. The standard theoretical account invokes spreading activation: when a cue is linked to multiple memory traces, activation spreads across them, reducing the target's retrievability. We study whether this spreading activation is triggered only by explicitly learned associations or also by...
Find it
Liu et al. (2023)
Evaluating Verifiability in Generative Search Engines
In
Findings of the Association for Computational Linguistics: EMNLP 2023, 7001-7025
Find it
Liu & Gabriel (2026)
PM-Bench: Evaluating Prospective Memory in LLM Agents
In
arXiv
Abstract
A significant challenge in agentic AI is prospective memory: the ability to execute an intention at a specific future cue or state while other activities are ongoing. We introduce PM-Bench, a text-based benchmark for measuring prospective memory capabilities in modern LLM agents. Inspired by the Virtual Week paradigm from cognitive science, PM-Bench evaluates how well LLM agents maintain user intentions, execute delayed intentions, and monitor latent environment changes. Over the course of a simulated seven-day week, agents must continue an ongoing activity while deciding whether any deferred task is due. We compare eight state-of-the-art LLMs on PM-Bench under eight different agent...
Find it
Lucas (2000)
Semantic priming without association: A meta-analytic review
In
Psychonomic Bulletin & Review, 7(4), 618-630
Find it
Luck & Vogel (1997)
The capacity of visual working memory for features and conjunctions
In
Nature, 390(6657), 279-281
Find it
Lum et al. (2014)
Procedural learning deficits in specific language impairment (SLI): A meta-analysis of serial reaction time task performance
In
Cortex, 51, 1-10
Find it
Luna et al. (2023)
Monitoring and control processes in mock witnesses in under-represented non-WEIRD samples with high or low educational level
In
Memory & Cognition, 51(3), 718-728
Find it
Ma & Xu (2025)
Implicit In-Context Learning: Evidence from Artificial Language Experiments
In
arXiv
In summary
Humans acquire language implicitly, absorbing complex patterns without explicit awareness, but whether LLMs show human-like pattern extraction during in-context learning at inference remains unclear. The authors adapt three classic artificial language learning experiments spanning morphology, morphosyntax, and syntax to evaluate implicit learning in gpt-4o and o3-mini. Results show alignment between models and human behaviour that is specific to linguistic domain: o3-mini aligns better in morphology, and both models align in syntax.
Find it
Maharana et al. (2024)
Evaluating Very Long-Term Conversational Memory of LLM Agents
In
Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 13851-13870
Abstract
Paraphrase, not the publisher's wording. Prior work on long-term open-domain dialogue evaluates model responses over no more than five chat sessions, leaving the effectiveness of long-context models and retrieval augmentation in genuinely long-term conversation untested. The authors introduce a machine-human pipeline that generates high-quality very long-term dialogues by grounding LLM agent conversations in personas and temporal event graphs, with agents able to share and react to images. Human annotators verify and edit the output for long-range consistency and grounding. The resulting dataset, LoCoMo, supports an evaluation benchmark covering question answering, event summarisation,...
Find it
Mahowald et al. (2016)
A meta-analysis of syntactic priming in language production
In
Journal of Memory and Language, 91, 5-27
Abstract
We performed an exhaustive meta-analysis of 73 peer-reviewed journal articles on syntactic priming from the seminal Bock (1986) paper through 2013. Extracting the effect size for each experiment and condition, where the effect size is the log odds ratio of the frequency of the primed structure X to the frequency of the unprimed structure Y, we found a robust effect of syntactic priming with an average weighted odds ratio of 1.67 when there is no lexical overlap and 3.26 when there is. That is, a construction X which occurs 50% of the time in the absence of priming would occur 63% if primed without lexical repetition and 77% of the time if primed with lexical repetition. The syntactic...
Find it
Mallen et al. (2023)
When Not to Trust Language Models: Investigating Effectiveness of Parametric and Non-Parametric Memories
In
Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 9802-9822
Abstract
Despite their impressive performance on diverse tasks, large language models (LMs) still struggle with tasks requiring rich world knowledge, implying the difficulty of encoding a wealth of world knowledge in their parameters. This paper aims to understand LMs’ strengths and limitations in memorizing factual knowledge, by conducting large-scale knowledge probing experiments on two open-domain entity-centric QA datasets: PopQA, our new dataset with 14k questions about long-tail entities, and EntityQuestions, a widely used open-domain QA dataset. We find that LMs struggle with less popular factual knowledge, and that retrieval augmentation helps significantly in these cases. Scaling, on the...
Find it
Mandler (1980)
Recognizing: The judgment of previous occurrence.
In
Psychological Review, 87(3), 252-271
Find it
Manning et al. (2011)
Oscillatory patterns in temporal lobe reveal context reinstatement during memory search
In
Proceedings of the National Academy of Sciences, 108(31), 12893-12897
Abstract
Psychological theories of memory posit that when people recall a past event, they not only recover the features of the event itself, but also recover information associated with other events that occurred nearby in time. The events surrounding a target event, and the thoughts they evoke, may be considered to represent a context for the target event, helping to distinguish that event from similar events experienced at different times. The ability to reinstate this contextual information during memory search has been considered a hallmark of episodic, or event-based, memory. We sought to determine whether context reinstatement may be observed in electrical signals recorded from the human...
Find it
Mansfield et al. (2026)
What might we learn about autobiographical narrative processing from Artificial Intelligence?
In
Humanities and Social Sciences Communications, 13(1)
Find it
Maren et al. (2013)
The contextual brain: implications for fear conditioning, extinction and psychopathology
In
Nature Reviews Neuroscience, 14(6), 417-428
Find it
Maril et al. (2003)
Feeling-of-knowing in episodic memory: an event-related fMRI study
In
NeuroImage, 18(4), 827-836
Find it
Martín‐Luengo et al. (2026)
External Source Monitoring and Reality Monitoring: Meta‐Analysis of <scp>fMRI</scp> Studies
In
Human Brain Mapping, 47(11)
Abstract
Being able to remember the origin of a particular memory is for times even more important than the memory itself. Yet, although memory for the information and memory for the source are intrinsically connected, the particular neural underpinnings for the source-memory monitoring remains unclear. We run an fMRI meta-analysis to help to clarify the brain structures necessary for source memory monitoring. Source monitoring (SM) relates to distinguishing the origin of the information of two external sources, whereas reality monitoring (RM) refers to distinguishing between one internal (e.g., thoughts) and one external source, or two internal sources. We used PRISMA to select the manuscripts...
Find it
McCall & Carriger (1993)
A Meta-Analysis of Infant Habituation and Recognition Memory Performance as Predictors of Later IQ
In
Child Development, 64(1), 57
Find it
McDaniel & Einstein (2000)
Strategic and automatic processes in prospective memory retrieval: a multiprocess framework
In
Applied Cognitive Psychology, 14(7)
Abstract
Prospective memory situations involve forming intentions and then realizing those intentions at some appropriate time in the future. An interesting feature of most prospective remembering is that recollection of the intended action occurs without an explicit request to attempt retrieval, and we present two views on how this type of remembering can be accomplished. One could strategically monitor the environment for the presence of the target event, or one could rely on anticipated environmental conditions more or less automatically reinstating the intended action. We present data supporting both views and argue that people use multiple approaches for solving the problem of retrieving an...
Find it
McDaniel et al. (2013)
Dissociable Neural Routes to Successful Prospective Memory
In
Psychological Science, 24(9), 1791-1800
Abstract
Identifying the processes by which people remember to execute an intention at an appropriate moment ( prospective memory ) remains a fundamental theoretical challenge. According to one account, top-down attentional control is required to maintain activation of the intention, initiate intention retrieval, or support monitoring. A diverging account suggests that bottom-up, spontaneous retrieval can be triggered by cues that have been associated with the intention and that sustained attentional processes are not required. We used a specialized experimental design and functional MRI methods to selectively marshal and identify each process. Results revealed a clear dissociation. One...
Find it
McGaugh (2000)
Memory--a Century of Consolidation
In
Science, 287(5451), 248-251
Abstract
The memory consolidation hypothesis proposed 100 years ago by Müller and Pilzecker continues to guide memory research. The hypothesis that new memories consolidate slowly over time has stimulated studies revealing the hormonal and neural influences regulating memory consolidation, as well as molecular and cellular mechanisms. This review examines the progress made over the century in understanding the time-dependent processes that create our lasting memories.
Find it
McGlinchey-Berroth et al. (1997)
Impaired trace eyeblink conditioning in bilateral, medial-temporal lobe amnesia.
In
Behavioral Neuroscience, 111(5), 873-882
Find it
Memon et al. (2010)
The Cognitive Interview: A meta-analytic review and study space analysis of the past 25 years.
In
Psychology, Public Policy, and Law, 16(4), 340-372
Find it
Mertens & Engelhard (2020)
A systematic review and meta-analysis of the evidence for unaware fear conditioning
In
Neuroscience & Biobehavioral Reviews, 108, 254-268
Find it
Metcalfe & Kornell (2003)
The Dynamics of Learning and Allocation of Study Time to a Region of Proximal Learning.
In
Journal of Experimental Psychology: General, 132(4), 530-542
Find it
Metcalfe & Finn (2008)
Evidence that judgments of learning are causally related to study choice
In
Psychonomic Bulletin & Review, 15(1), 174-179
Find it
Meunier-Duperray et al. (2025)
On the complexity of metacognitive judgments of memory: evidence from retrospective confidence, feeling of knowing, and older adults
In
Neuroscience of Consciousness, 2025(1)
Abstract
Abstract Dissociations in types of memory tasks emerge when comparing feeling-of-knowing (FOK) judgments, predictions of upcoming performance, and retrospective confidence. This pattern has been used to construct theories of metacognitive access to memory, particularly in memory-impaired groups. In particular, older adults’ metacognitive sensitivity appears to vary between episodic (impaired) and semantic (intact) memory. However, this could be explained by the limitations of metacognitive measures and/or memory differences. We aimed to test these dissociations of metacognition with aging by comparing metacognitive efficiency in episodic and semantic tasks using two types of judgment:...
Find it
Meyer & Benoit (2022)
Suppression weakens unwanted memories via a sustained reduction of neural reactivation
In
eLife, 11
Abstract
Aversive events sometimes turn into intrusive memories. However, prior evidence indicates that such memories can be controlled via a mechanism of retrieval suppression. Here, we test the hypothesis that suppression exerts a sustained influence on memories by deteriorating their neural representations. This deterioration, in turn, would hinder their subsequent reactivation and thus impoverish the vividness with which they can be recalled. In an fMRI study, participants repeatedly suppressed memories of aversive scenes. As predicted, this process rendered the memories less vivid. Using a pattern classifier, we observed that suppression diminished the neural reactivation of scene information...
Find it
Michaelov et al. (2023)
Structural Priming Demonstrates Abstract Grammatical Representations in Multilingual Language Models
In
Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, 3703-3720
Find it
Migues et al. (2016)
Blocking Synaptic Removal of GluA2-Containing AMPA Receptors Prevents the Natural Forgetting of Long-Term Memories
In
The Journal of Neuroscience, 36(12), 3481-3494
Abstract
The neurobiological processes underpinning the natural forgetting of long-term memories are poorly understood. Based on the critical role of GluA2-containing AMPA receptors (GluA2/AMPARs) in long-term memory persistence, we tested in rats whether their synaptic removal underpins time-dependent memory loss. We found that blocking GluA2/AMPAR removal with the interference peptides GluA2 3Y or G2CT in the dorsal hippocampus during a memory retention interval prevented the normal forgetting of established, long-term object location memories, but did not affect their acquisition. The same intervention also preserved associative memories of food-reward conditioned place preference that would...
Find it
Mireshghallah, Niloofar, Neal Mangaokar, Narine Kokhlikyan, Arman Zharmagambetov, Manzil Zaheer, Saeed Mahloujifar, and Kamalika Chaudhuri (2026)
CIMemories: A Compositional Benchmark For Contextual Integrity In LLMs
In
International Conference on Learning Representations, 2026, 95513-95532
Abstract
Large Language Models (LLMs) increasingly use persistent memory from past interactions to enhance personalization and task performance. However, this memory introduces critical risks when sensitive information is revealed in inappropriate contexts. We present CIMemories, a benchmark for evaluating whether LLMs appropriately control information flow from memory based on task context. CIMemories uses synthetic user profiles with over 100 attributes per user, paired with diverse task contexts in which each attribute may be essential for some tasks but inappropriate for others. Our evaluation reveals that frontier models exhibit up to 69% attribute-level violations (leaking information...
Find it
Mitchell (2006)
Nonconscious Priming After 17 Years
In
Psychological Science, 17(11), 925-929
Abstract
Individuals who saw pictures for 1 to 3 s in the laboratory were tested 17 years later by mail. Identification rates were significantly higher for fragments from these previously exposed targets than for novel fragments, whereas the same stimuli evoked no differences in control groups that had not been previously exposed to the pictures. Priming—the memorial advantage conferred by prior perceptual experience—was stable over the years (r = .51). Priming was dissociated from episodic memory, in that it was present even in subjects who reported no conscious recollection of their participation in the original laboratory session. These findings suggest that the perceptual representation system...
Find it
Mittal (2026)
Did You Forget What I Asked? Prospective Memory Failures in Large Language Models
In
Proceedings of the 6th Workshop on Trustworthy NLP (TrustNLP 2026), 471-488
Abstract
Large language models often fail to satisfy formatting instructions when they must simultaneously perform demanding tasks. We study this behavior through a prospective memory-inspired lens from cognitive psychology, using a controlled paradigm that combines verifiable formatting constraints with benchmark tasks of increasing complexity. Across three model families and over 8,000 prompts, compliance drops by 2–21% under concurrent task load. Vulnerability is highly type-dependent: terminal constraints (requiring action at the response boundary) degrade most, with drops up to 50%, while avoidance constraints remain comparatively robust. A salience-enhanced format (explicit instruction...
Find it
Miyamoto et al. (2017)
Causal neural network of metamemory for retrospection in primates
In
Science, 355(6321), 188-193
Abstract
Are you aware how well you remember? Self-monitoring and evaluation of our own memory is a mental process called metamemory. For metamemory, we need access to information about the strength of our own memory traces. The brain structures and neural mechanisms involved in metamemory are completely unknown. Miyamoto et al. devised a test paradigm for metamemory in macaques, in which the monkeys judged their own confidence in remembering past experiences. The authors combined this approach with functional brain imaging to reveal the neural substrates of metamemory for retrospection. A specific region in the prefrontal brain was essential for meta mnemonic decision-making. Inactivation of this...
Find it
Modarressi et al. (2025)
NoLiMa: Long-Context Evaluation Beyond Literal Matching
In
arXiv
Abstract
Recent large language models (LLMs) support long contexts ranging from 128K to 1M tokens. A popular method for evaluating these capabilities is the needle-in-a-haystack (NIAH) test, which involves retrieving a "needle" (relevant information) from a "haystack" (long irrelevant context). Extensions of this approach include increasing distractors, fact chaining, and in-context reasoning. However, in these benchmarks, models can exploit existing literal matches between the needle and haystack to simplify the task. To address this, we introduce NoLiMa, a benchmark extending NIAH with a carefully designed needle set, where questions and needles have minimal lexical overlap, requiring models to...
Find it
Modirrousta & Fellows (2008)
Medial prefrontal cortex plays a critical and selective role in ‘feeling of knowing’ meta-memory judgments
In
Neuropsychologia, 46(12), 2958-2965
Find it
Munawar et al. (2018)
Understanding the reminiscence bump: A systematic review
In
PLOS ONE, 13(12), e0208595
Find it
Murre & Dros (2015)
Replication and Analysis of Ebbinghaus’ Forgetting Curve
In
PLOS ONE, 10(7), e0120644
Find it
Murre (2021)
The Godden and Baddeley (1975) experiment on context-dependent memory on land and underwater: a replication
In
Royal Society Open Science, 8(11)
Abstract
A replication of the experiment by Godden and Baddeley (Godden and Baddeley 1975 British Journal of Psychology 66, 325–331 (doi:10.1111/j.2044-8295.1975.tb01468.x)) on environmental context-dependent memory is described. Sixteen divers studied auditorily presented word lists on land or underwater and recalled these 4 min later on land or underwater (each diver participated in all four combinations). Contrary to the original study, we did not find that recall in the same context where the words had been learned was better than recall in the other context. We discuss differences between our replication and the original study and emphasize the importance of attempts at replicating classic...
Find it
Nader et al. (2000)
Fear memories require protein synthesis in the amygdala for reconsolidation after retrieval
In
Nature, 406(6797), 722-726
Abstract
‘New’ memories are initially labile and sensitive to disruption before being consolidated into stable long-term memories1,2,3,4,5. Much evidence indicates that this consolidation involves the synthesis of new proteins in neurons6,7,8,9. The lateral and basal nuclei of the amygdala (LBA) are believed to be a site of memory storage in fear learning10. Infusion of the protein synthesis inhibitor anisomycin into the LBA shortly after training prevents consolidation of fear memories11. Here we show that consolidated fear memories, when reactivated during retrieval, return to a labile state in which infusion of anisomycin shortly after memory reactivation produces amnesia on later tests,...
Find it
Nader & Hardt (2009)
A single standard for memory: the case for reconsolidation
In
Nature Reviews Neuroscience, 10(3), 224-234
Find it
Nault et al. (2023)
Investigating the influence of local and personal common ground on memory for conversation using an online referential communication task.
In
Journal of Experimental Psychology: General, 152(6), 1598-1621
Find it
Nelson (1990)
Metamemory: A Theoretical Framework and New Findings
In
Psychology of Learning and Motivation, 125-173
Find it
Newbury et al. (2021)
Sleep deprivation and memory: Meta-analytic reviews of studies on sleep deprivation before and after learning.
In
Psychological Bulletin, 147(11), 1215-1240
Find it
Nissen & Bullemer (1987)
Attentional requirements of learning: Evidence from performance measures
In
Cognitive Psychology, 19(1), 1-32
Find it
Old & Naveh-Benjamin (2008)
Differential effects of age on item and associative measures of memory: A meta-analysis.
In
Psychology and Aging, 23(1), 104-118
Find it
Oliveira et al. (2023)
The reliability of the serial reaction time task: meta-analysis of test–retest correlations
In
Royal Society Open Science, 10(7)
Abstract
The Serial Reaction Time task, one of the most widely used tasks to index procedural memory, has been increasingly employed in individual differences research examining the role of procedural memory in language and other cognitive abilities. Yet, despite consistently producing robust procedural learning effects at the group level (i.e. faster responses to sequenced/probable trials versus random/improbable trials), these effects have recently been found to have poor reliability. In this meta-analysis ( N = 7), comprising 719 participants ( M = 20.81, s.d. = 7.13), we confirm this ‘reliability paradox’. The overall retest reliability of the robust procedural learning effect elicited by the...
Find it
Orban et al. (2022)
Verbal Memory Interference in Attention-Deficit Hyperactivity Disorder: A Meta-Analytic Review
In
Journal of Attention Disorders, 26(12), 1549-1562
Abstract
Objective: Interference control is used to overcome conflict among competing memory representations and may contribute to memory difficulties in ADHD. This meta-analytic review examined memory interference to evaluate susceptibility to proactive, retroactive, and memory control interference among those with ADHD. Method: Twenty studies (1987–2019) examining verbal memory interference in ADHD met inclusion criteria (age: 8–36 years). Proactive and retroactive interference indices were extracted from list-learning tasks, and memory control indices were extracted from experimental paradigms (e.g., directed-forgetting). Results: Children with ADHD were less affected by proactive interference...
Find it
Orhan (2023)
Recognition, recall, and retention of few-shot memories in large language models
In
arXiv
Abstract
The training of modern large language models (LLMs) takes place in a regime where most training examples are seen only a few times by the model during the course of training. What does a model remember about such examples seen only a few times during training and how long does that memory persist in the face of continuous training with new examples? Here, we investigate these questions through simple recognition, recall, and retention experiments with LLMs. In recognition experiments, we ask if the model can distinguish the seen example from a novel example; in recall experiments, we ask if the model can correctly recall the seen example when cued by a part of it; and in retention...
Find it
Ouyang et al. (2022)
Training language models to follow instructions with human feedback
In
arXiv
Abstract
Making language models bigger does not inherently make them better at following a user's intent. For example, large language models can generate outputs that are untruthful, toxic, or simply not helpful to the user. In other words, these models are not aligned with their users. In this paper, we show an avenue for aligning language models with user intent on a wide range of tasks by fine-tuning with human feedback. Starting with a set of labeler-written prompts and prompts submitted through the OpenAI API, we collect a dataset of labeler demonstrations of the desired model behavior, which we use to fine-tune GPT-3 using supervised learning. We then collect a dataset of rankings of model...
Find it
Owen et al. (2005)
N‐back working memory paradigm: A meta‐analysis of normative functional neuroimaging studies
In
Human Brain Mapping, 25(1), 46-59
Abstract
Abstract One of the most popular experimental paradigms for functional neuroimaging studies of working memory has been the n‐back task, in which subjects are asked to monitor the identity or location of a series of verbal or nonverbal stimuli and to indicate when the currently presented stimulus is the same as the one presented n trials previously. We conducted a quantitative meta‐analysis of 668 sets of activation coordinates in Talairach space reported in 24 primary studies of n‐back task variants manipulating process (location vs. identity monitoring) and content (verbal or nonverbal) of working memory. We found the following cortical regions were activated robustly (voxelwise false...
Find it
Pachur et al. (2014)
We'll Meet Again: Revealing Distributional and Temporal Patterns of Social Contact
In
PLoS ONE, 9(1), e86081
Find it
Packard & Knowlton (2002)
Learning and Memory Functions of the Basal Ganglia
In
Annual Review of Neuroscience, 25(1), 563-593
Abstract
▪ Abstract Although the mammalian basal ganglia have long been implicated in motor behavior, it is generally recognized that the behavioral functions of this subcortical group of structures are not exclusively motoric in nature. Extensive evidence now indicates a role for the basal ganglia, in particular the dorsal striatum, in learning and memory. One prominent hypothesis is that this brain region mediates a form of learning in which stimulus-response (S-R) associations or habits are incrementally acquired. Support for this hypothesis is provided by numerous neurobehavioral studies in different mammalian species, including rats, monkeys, and humans. In rats and monkeys, localized brain...
Find it
Pan & Rickard (2015)
Sleep and motor learning: Is there room for consolidation?
In
Psychological Bulletin, 141(4), 812-834
Find it
Panickssery et al. (2024)
LLM Evaluators Recognize and Favor Their Own Generations
In
Advances in Neural Information Processing Systems 37, 68772-68802
Find it
Pannu & Kaszniak (2005)
Metamemory Experiments in Neurological Populations: A Review
In
Neuropsychology Review, 15(3), 105-130
Find it
Park et al. (2023)
Generative Agents: Interactive Simulacra of Human Behavior
In
Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology, 1-22
Abstract
Believable proxies of human behavior can empower interactive applications ranging from immersive environments to rehearsal spaces for interpersonal communication to prototyping tools. In this paper, we introduce generative agents: computational software agents that simulate believable human behavior. Generative agents wake up, cook breakfast, and head to work; artists paint, while authors write; they form opinions, notice each other, and initiate conversations; they remember and reflect on days past as they plan the next day. To enable generative agents, we describe an architecture that extends a large language model to store a complete record of the agent’s experiences using natural...
Find it
Peng & Tullis (2021)
Dividing attention impairs metacognitive control more than monitoring
In
Psychonomic Bulletin & Review, 28(6), 2064-2074
Find it
Pereira et al. (2023)
Examining the effects of emotional valence and arousal on source memory: A meta-analysis of behavioral evidence.
In
Emotion, 23(6), 1740-1763
Find it
Pessiglione et al. (2006)
Dopamine-dependent prediction errors underpin reward-seeking behaviour in humans
In
Nature, 442(7106), 1042-1045
Find it
Peter et al. (2026)
The functional neuroanatomy of event-based and time-based prospective memory, and its improvement
In
Communications Biology, 9(1)
Find it
Pevie et al. (2023)
Item-method directed forgetting is (usually) impaired in clinical populations: A meta-analysis.
In
Canadian Journal of Experimental Psychology / Revue canadienne de psychologie expérimentale, 77(4), 271-283
Find it
Pickering & Ferreira (2008)
Structural priming: A critical review.
In
Psychological Bulletin, 134(3), 427-459
Find it
Pigeon et al. (2022)
Impairing memory reconsolidation with propranolol in healthy and clinical samples: a meta-analysis
In
Journal of Psychiatry and Neuroscience, 47(2), E109-E122
Abstract
Background Reconsolidation impairment using propranolol is a novel intervention for mental disorders with an emotional memory at their core. In this systematic review and meta-analysis, we examined the evidence for this intervention in healthy and clinical adult samples. Methods We searched 8 databases for randomized, double-blind studies that involved at least 1 propranolol group and 1 placebo group. We conducted a meta-analysis of 14 studies ( n = 478) in healthy adults and 12 studies in clinical samples ( n = 446). Results Compared to placebo, reconsolidation impairment under propranolol resulted in reduced recall of aversive material and cue-elicited conditioned emotional responses in...
Find it
Pink et al. (2024)
Assessing Episodic Memory in LLMs with Sequence Order Recall Tasks
In
arXiv
Abstract
Current LLM benchmarks focus on evaluating models' memory of facts and semantic relations, primarily assessing semantic aspects of long-term memory. However, in humans, long-term memory also includes episodic memory, which links memories to their contexts, such as the time and place they occurred. The ability to contextualize memories is crucial for many cognitive tasks and everyday functions. This form of memory has not been evaluated in LLMs with existing benchmarks. To address the gap in evaluating memory in LLMs, we introduce Sequence Order Recall Tasks (SORT), which we adapt from tasks used to study episodic memory in cognitive psychology. SORT requires LLMs to recall the correct...
Find it
Pink et al. (2026)
Temporal Context Reinstatement Drives Episodic-Like Order Memory in Long-Context Language Models
In
arXiv
Abstract
Human episodic memory supports the retrieval of experiences that unfold over extended timescales, yet the computational mechanisms underlying this ability remain debated due to the limited mechanistic accessibility in long-term memory experiments in humans. Long-context LLMs may offer promising ways to reveal plausible computational mechanisms that drive this type of retrieval. Here, we investigate whether and how LLMs capture the core behavioral signatures of episodic memory via a temporal order memory task. Using a new dataset of human behavior based on memory of a full-length novel, we show that models exhibit the same characteristic distance effect observed in humans on this task. We...
Find it
Plichta et al. (2014)
Amygdala habituation: A reliable fMRI phenotype
In
NeuroImage, 103, 383-390
Find it
Pole (2007)
The psychophysiology of posttraumatic stress disorder: A meta-analysis.
In
Psychological Bulletin, 133(5), 725-746
Find it
Polyn et al. (2005)
Category-Specific Cortical Activity Precedes Retrieval During Memory Search
In
Science, 310(5756), 1963-1966
Abstract
Here we describe a functional magnetic resonance imaging study of humans engaged in memory search during a free recall task. Patterns of cortical activity associated with the study of three categories of pictures (faces, locations, and objects) were identified by a pattern-classification algorithm. The algorithm was used to track the reappearance of these activity patterns during the recall period. The reappearance of a given category's activity pattern correlates with verbal recalls made from that category and precedes the recall event by several seconds. This result is consistent with the hypothesis that category-specific activity is cueing the memory system to retrieve studied items.
Find it
Posner & Keele (1968)
On the genesis of abstract ideas.
In
Journal of Experimental Psychology, 77(3, Pt.1), 353-363
Find it
Qiao (2025)
Attention as Conditioning: What Classical Learning Theory Predicts About Linear Transformers
In
arXiv
Abstract
Attention is widely understood as an associative memory, but that description alone does not predict how the memory will behave. Predictive theories do exist, but in the literature on animal learning. We show that the state updates of the major linear-attention families are term-for-term identical with named models from a century of animal learning theory: linear attention implements Hebbian contiguity, DeltaNet implements Rescorla--Wagner error correction, and decay variants such as RetNet implement contiguity with a stimulus trace. This dictionary turns conditioning phenomena into testable statements about the in-context behavior of linear transformers, while distinguishing algebraic...
Find it
Qin et al. (2026)
ImplicitMemBench: Measuring Unconscious Behavioral Adaptation in Large Language Models
In
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 28232-28261
Abstract
Existing memory benchmarks for LLM agents evaluate explicit recall of facts, yet overlook implicit memory where experience becomes automated behavior without conscious retrieval. This gap is critical: effective assistants must automatically apply learned procedures or avoid failed actions without explicit reminders. We introduce ImplicitMemBench, the first systematic benchmark evaluating implicit memory through three cognitively grounded constructs drawn from standard cognitive-science accounts of non-declarative memory: Procedural Memory (one-shot skill acquisition after interference), Priming (theme-driven bias via paired experimental/control instances), and Classical Conditioning...
Find it
Ramesh (2026)
Can an AI Assistant Really Forget? Auditable Deletion from Addressable Memory
In
arXiv
Abstract
An assistant can stop repeating a fact without removing it from memory. To study this difference, we install a support-vector gate in frozen Gemma 3 and record which stored keys and values belong to each exchange. A deletion request excludes the exchange's rows from the long-range readout and recalculates the gate on what remains. We check this operation against an independent refit, then compare it with running the model again on the conversation without the exchange. This second comparison matters because the exchange may already have influenced surviving memory rows. At 4B, the gated model passed checks for recall and feasible deletion on the same six of eight records admitted by the...
Find it
Ranjan & Odegaard (2024)
Reality monitoring and metacognitive judgments in a false-memory paradigm
In
Neuroscience Research, 201, 3-17
Abstract
How well do we distinguish between different memory sources when the information from imagination and perception is similar? And how do metacognitive (confidence) judgments differ across different sources of experiences? To study these questions, we developed a reality monitoring task using semantically related words from the Deese-Roediger-McDermott (DRM) paradigm of false memories. In an orientation phase, participants either perceived word pairs or had to voluntarily imagine the second word of a word pair. In a test phase, participants viewed words and had to judge whether the paired word was previously perceived, imagined, or new. Results revealed an interaction between memory source...
Find it
Ranjan et al. (2026)
Reality Monitoring in Large Language Models: Self-Knowledge That Transforms with Conversation Memory
In
arXiv
Abstract
A conversational AI that cannot tell its own output from what a user said will treat its own mistakes as user-provided facts. In humans, this capacity is called reality monitoring, and its failures are linked to hallucinations, delusions, and confabulation, yet whether LLMs possess it remains untested. Here we show, across two experiments and six LLMs, that source attribution depends on how conversational memory is structured: ceiling accuracy for self-generated content under minimal memory demands reverses to a fragile external-item advantage once episodic delay removes that shortcut. Feedback exposes two failures: in some models, internal and external judgments swap; in others, accuracy...
Find it
Rankin et al. (2009)
Habituation revisited: An updated and revised description of the behavioral characteristics of habituation
In
Neurobiology of Learning and Memory, 92(2), 135-138
Abstract
The most commonly cited descriptions of the behavioral characteristics of habituation come from two papers published almost 40 years ago (Thompson and Spencer, 1966; Groves and Thompson, 1970). In August 2007, the authors of this review, who study habituation in a wide range of species and paradigms, met to discuss their work on habituation and to revisit and refine the characteristics of habituation. This review offers a re-evaluation of the characteristics of habituation in light of these discussions. We made substantial changes to only a few of the characteristics, usually to add new information and expand upon the description rather than to substantially alter the original point.
Find it
Rasheed & Kabtoul (2026)
eMEM: A Hybrid Spatio-Temporal Memory System For Embodied Agents
In
arXiv
Abstract
We present eMEM (Embodied Memory), a hybrid graph-based memory system for embodied agents operating in physical environments. Current agent memory architectures, such as Generative Agents, MemGPT, and A-MEM, treat memory as text streams or knowledge graphs, but embodied agents require memory that is simultaneously searchable by meaning, space, and time. eMEM fills this gap with a multi-index architecture (SQLITE for structured storage, hnswlib for approximate nearest neighbour semantic search, and an R-tree for spatial queries) unified behind a single graph model. A tiered consolidation pipeline transforms raw perceptual observations into compressed summaries, mirroring...
Find it
Rasmussen & Berntsen (2011)
The unpredictable past: Spontaneous autobiographical memories outnumber autobiographical memories retrieved strategically
In
Consciousness and Cognition, 20(4), 1842-1846
Find it
Reber (1967)
Implicit learning of artificial grammars
In
Journal of Verbal Learning and Verbal Behavior, 6(6), 855-863
Find it
Reder & Ritter (1992)
What determines initial feeling of knowing? Familiarity with question terms, not with the answer.
In
Journal of Experimental Psychology: Learning, Memory, and Cognition, 18(3), 435-451
Find it
Redick & Lindsey (2013)
Complex span and n-back measures of working memory: A meta-analysis
In
Psychonomic Bulletin & Review, 20(6), 1102-1113
Find it
Renoult et al. (2012)
Personal semantics: at the crossroads of semantic and episodic memory
In
Trends in Cognitive Sciences, 16(11), 550-558
Find it
Renoult et al. (2019)
From Knowing to Remembering: The Semantic–Episodic Distinction
In
Trends in Cognitive Sciences, 23(12), 1041-1057
Find it
Rescorla (1968)
Probability of shock in the presence and absence of cs in fear conditioning.
In
Journal of Comparative and Physiological Psychology, 66(1), 1-5
Find it
Rhodes & Tauber (2011)
The influence of delaying judgments of learning on metacognitive accuracy: A meta-analytic review.
In
Psychological Bulletin, 137(1), 131-148
Find it
Rhodes et al. (2019)
Age-related differences in recall and recognition: a meta-analysis
In
Psychonomic Bulletin & Review, 26(5), 1529-1547
Find it
Richards & Frankland (2017)
The Persistence and Transience of Memory
In
Neuron, 94(6), 1071-1084
Abstract
The predominant focus in the neurobiological study of memory has been on remembering (persistence). However, recent studies have considered the neurobiology of forgetting (transience). Here we draw parallels between neurobiological and computational mechanisms underlying transience. We propose that it is the interaction between persistence and transience that allows for intelligent decision-making in dynamic, noisy environments. Specifically, we argue that transience (1) enhances flexibility, by reducing the influence of outdated information on memory-guided decision-making, and (2) prevents overfitting to specific past events, thereby promoting generalization. According to this view, the...
Find it
Rojahn & Pettigrew (1992)
Memory for schema‐relevant information: A meta‐analytic resolution
In
British Journal of Social Psychology, 31(2), 81-109
Abstract
The research on memory for schema‐relevant information has produced sharply contradictory results. Some studies demonstrate memory selectivity for schema‐consistent information, others demonstrate memory selectivity for schema‐inconsistent information. Meta‐analytic procedures are applied to a sample of 60 independent studies with 165 comparative tests. The overall result shows a slight overall memory advantage for schema‐ inconsistent information. As the effect sizes are highly heterogeneous, it was hypothesized that schema‐based processing is moderated by an array of variables. In large part the direction of the effect is dependent on the memory measure used. Both recognition tests...
Find it
Rönnlund et al. (2005)
Stability, Growth, and Decline in Adult Life Span Development of Declarative Memory: Cross-Sectional and Longitudinal Data From a Population-Based Study.
In
Psychology and Aging, 20(1), 3-18
Find it
Rosenbaum et al. (2005)
The case of K.C.: contributions of a memory-impaired person to memory theory
In
Neuropsychologia, 43(7), 989-1021
Find it
Rubin & Baddeley (1989)
Telescoping is not time compression: A model
In
Memory & Cognition, 17(6), 653-661
Find it
Rubin & Wenzel (1996)
One hundred years of forgetting: A quantitative description of retention.
In
Psychological Review, 103(4), 734-760
Find it
Rubin & Schulkind (1997)
The distribution of autobiographical memories across the lifespan
In
Memory & Cognition, 25(6), 859-866
Find it
Rubin et al. (2011)
How Do I Remember That I Know You Know That I Know?
In
Psychological Science, 22(12), 1574-1582
Abstract
Communication is aided greatly when speakers and listeners take advantage of mutually shared knowledge (i.e., common ground). How such information is represented in memory is not well known. Using a neuropsychological-psycholinguistic approach to real-time language understanding, we investigated the ability to form and use common ground during conversation in memory-impaired participants with hippocampal amnesia. Analyses of amnesics’ eye fixations as they interpreted their partner’s utterances about a set of objects demonstrated successful use of common ground when the amnesics had immediate access to common-ground information, but dramatic failures when they did not. These findings...
Find it
Schacter et al. (2007)
Reductions in cortical activity during priming
In
Current Opinion in Neurobiology, 17(2), 171-176
Find it
Schlagman & Kvavilashvili (2008)
Involuntary autobiographical memories in and outside the laboratory: How different are they from voluntary autobiographical memories?
In
Memory & Cognition, 36(5), 920-932
Find it
Schmitz et al. (2017)
Hippocampal GABA enables inhibitory control over unwanted thoughts
In
Nature Communications, 8(1)
Abstract
AbstractIntrusive memories, images, and hallucinations are hallmark symptoms of psychiatric disorders. Although often attributed to deficient inhibitory control by the prefrontal cortex, difficulty in controlling intrusive thoughts is also associated with hippocampal hyperactivity, arising from dysfunctional GABAergic interneurons. How hippocampal GABA contributes to stopping unwanted thoughts is unknown. Here we show that GABAergic inhibition of hippocampal retrieval activity forms a key link in a fronto-hippocampal inhibitory control pathway underlying thought suppression. Subjects viewed reminders of unwanted thoughts and tried to suppress retrieval while being scanned with functional...
Find it
Schneider & Anderson (2012)
Modeling fan effects on the time course of associative recognition
In
Cognitive Psychology, 64(3), 127-160
Find it
Schnyer et al. (2004)
A role for right medial prefrontal cortex in accurate feeling-of-knowing judgments: evidence from patients with lesions to frontal cortex
In
Neuropsychologia, 42(7), 957-966
Abstract
The hypothesis that prefrontal cortex plays a critical role in accurate predictions of episodic memory performance was tested using the feeling-of-knowing (FOK) paradigm. Fourteen patients with a broad spectrum of damage to the frontal cortex and matched controls read sentences and later were tested for recall memory, confidence judgments, and FOK accuracy using as cues the sentences with the final word missing. While frontal patients were impaired at recall and recognition memory, they were able to make accurate confidence judgments about their recall attempts. By contrast, as a group, the patients were markedly impaired in the accuracy of their prospective FOK judgments. Lesion analysis...
Find it
Schug & Lake (2026)
Are they human? Detecting large language models by probing human memory constraints
In
arXiv
Abstract
The validity of online behavioral research relies on study participants being human rather than machine. In the past, it was possible to detect machines by posing simple challenges that were easily solved by humans but not by machines. General-purpose agents based on large language models (LLMs) can now solve many of these challenges, threatening the validity of online behavioral research. Here we explore the idea of detecting humanness by using tasks that machines can solve too well to be human. Specifically, we probe for the existence of an established human cognitive constraint: limited working memory capacity. We show that cognitive modeling on a standard serial recall task can be...
Find it
Scoville & Milner (1957)
LOSS OF RECENT MEMORY AFTER BILATERAL HIPPOCAMPAL LESIONS
In
Journal of Neurology, Neurosurgery & Psychiatry, 20(1), 11-21
Find it
Scully et al. (2017)
Does reactivation trigger episodic memory change? A meta-analysis
In
Neurobiology of Learning and Memory, 142, 99-107
Find it
Segaert et al. (2012)
Shared Syntax in Language Production and Language Comprehension--An fMRI Study
In
Cerebral Cortex, 22(7), 1662-1670
Abstract
During speaking and listening syntactic processing is a crucial step. It involves specifying syntactic relations between words in a sentence. If the production and comprehension modality share the neuronal substrate for syntactic processing then processing syntax in one modality should lead to adaptation effects in the other modality. In the present functional magnetic resonance imaging experiment, participants either overtly produced or heard descriptions of pictures. We looked for brain regions showing adaptation effects to the repetition of syntactic structures. In order to ensure that not just the same brain regions but also the same neuronal populations within these regions are...
Find it
Sevenster et al. (2013)
Prediction Error Governs Pharmacologically Induced Amnesia for Learned Fear
In
Science, 339(6121), 830-833
Abstract
Overwriting Human Memory A consolidated fear memory can enter a transient labile phase upon its reactivation. Pharmacological blockade of the subsequent protein synthesis-dependent restabilization produces a memory deficit. Sevenster et al. (p. 830 ) found that prediction error (a discrepancy between actual and expected events) is a necessary condition for reconsolidation of a fear memory in human subjects. Retrieval cued by a negative prediction error, for example, like a conditioned stimulus presented in the absence of the unconditioned stimulus, resulted in labilization and subsequent reconsolidation was blocked by the drug propranolol. However, if there was no prediction error because...
Find it
Shalowitz et al. (2006)
The Accuracy of Surrogate Decision Makers
In
Archives of Internal Medicine, 166(5), 493
Find it
Shan et al. (2025)
Involuntary autobiographical memories as a transdiagnostic factor in mental disorders
In
Clinical Psychology Review, 116, 102545
Find it
Shepard (1967)
Recognition memory for words, sentences, and pictures
In
Journal of Verbal Learning and Verbal Behavior, 6(1), 156-163
Find it
Shimamura et al. (1990)
Memory for the temporal order of events in patients with frontal lobe lesions and amnesic patients
In
Neuropsychologia, 28(8), 803-813
Find it
Shinn et al. (2023)
Reflexion: Language Agents with Verbal Reinforcement Learning
In
arXiv
Abstract
Large language models (LLMs) have been increasingly used to interact with external environments (e.g., games, compilers, APIs) as goal-driven agents. However, it remains challenging for these language agents to quickly and efficiently learn from trial-and-error as traditional reinforcement learning methods require extensive training samples and expensive model fine-tuning. We propose Reflexion, a novel framework to reinforce language agents not by updating weights, but instead through linguistic feedback. Concretely, Reflexion agents verbally reflect on task feedback signals, then maintain their own reflective text in an episodic memory buffer to induce better decision-making in...
Find it
Simons et al. (2017)
Brain Mechanisms of Reality Monitoring
In
Trends in Cognitive Sciences, 21(6), 462-473
Abstract
Reality monitoring processes are necessary for discriminating between internally generated information and information that originated in the outside world. They help us to identify our thoughts, feelings, and imaginations, and to distinguish them from events we may have experienced or have been told about by someone else. Reality monitoring errors range from confusions between real and imagined experiences, that are byproducts of normal cognition, to symptoms of mental illness such as hallucinations. Recent advances support an emerging neurocognitive characterization of reality monitoring that provides insights into its underlying operating principles and neural mechanisms, the differing...
Find it
Sinclair et al. (2026)
Structural priming in humans and large language models
In
Journal of Memory and Language, 149, 104713
Find it
Singer & Tiede (2008)
Feeling of knowing and duration of unsuccessful memory search
In
Memory & Cognition, 36(3), 588-597
Find it
Smith & Vela (2001)
Environmental context-dependent memory: A review and meta-analysis
In
Psychonomic Bulletin & Review, 8(2), 203-220
Abstract
To address questions about human memory’s dependence on the coincidental environmental contexts in which events occur, we review studies of incidental environmental context-dependent memory in humans and report a meta-analysis. Our theoretical approach to the issue stems from Glenberg’s (1997) contention that introspective thought (e.g., remembering, conceptualizing) requires cognitive resources normally used to represent the immediate environment. We propose that if tasks encourage processing of noncontextual information (i.e., introspective thought) at input and/or at test, then both learning and memory will be less dependent on the ambient environmental contexts in which those...
Find it
Smith (2003)
The cost of remembering to remember in event-based prospective memory: Investigating the capacity demands of delayed intention performance.
In
Journal of Experimental Psychology: Learning, Memory, and Cognition, 29(3), 347-361
Find it
Smith et al. (2005)
Acquisition of Differential Delay Eyeblink Classical Conditioning Is Independent of Awareness.
In
Behavioral Neuroscience, 119(1), 78-86
Abstract
There has been debate about whether differential delay eyeblink conditioning can be acquired without awareness of the stimulus contingencies. In 4 experiments, the authors reexamined this question. Older participants were tested with a tone and white noise (Experiment 1) or with 2 tones (Experiment 2). In addition, younger participants were tested with 2 tones (Experiment 3) or with 2 tones plus the parameters from an earlier study that had reported a relationship between conditioning and awareness (Experiment 4). Participants who were designated aware of the stimulus contingencies and participants who were designated unaware exhibited equivalent levels of differential eyeblink...
Find it
Smith & Squire (2005)
Declarative Memory, Awareness, and Transitive Inference
In
The Journal of Neuroscience, 25(44), 10138-10146
Abstract
A characteristic usually attributed to declarative memory is that what is learned is accessible to awareness. Recently, the relationship between awareness and declarative (hippocampus-dependent) memory has been questioned on the basis of findings from transitive inference tasks. In transitive inference, participants are first trained on overlapping pairs of items (e.g., A+B-, B+C-, C+D-, and D+E-, where + and - indicate correct and incorrect choices). Later, participants who choose B over D when presented with the novel pair BD are said to demonstrate transitive inference. The ability to exhibit transitive inference is thought to depend on the fact that participants have represented the...
Find it
Sohn et al. (2003)
Competition and representation during memory retrieval: Roles of the prefrontal cortex and the posterior parietal cortex
In
Proceedings of the National Academy of Sciences, 100(12), 7412-7417
Abstract
In this functional-MRI study we examined the hypothesis that the prefrontal cortex responds differently to the extent of competition during retrieval, whereas the parietal cortex is responsible for problem representation that should not be directly related to the competition. Participants mastered arbitrary person–location pairs, and their recognition memory was tested in a functional-MRI session. The pairs were constructed such that a person was associated with one, two, or three different locations and vice versa. The recognition time increased with the number of associations, reflecting increased competition. A confirmatory analysis of imaging data with prespecified prefrontal and...
Find it
Spalding et al. (2015)
Investigating the Neural Correlates of Schemas: Ventromedial Prefrontal Cortex Is Necessary for Normal Schematic Influence on Memory
In
The Journal of Neuroscience, 35(47), 15746-15751
Abstract
Schemas, as memory representations of typical contexts, allow for generalization from previous experiences while often improving memory organization and accuracy. However, these advantageous characteristics of schematic memory may come at the cost of episode-specific information. In the human brain, this tradeoff between general and specific knowledge has been linked to differential contributions of the medial temporal lobes (MTL) to episode-specific memory and the ventromedial prefrontal cortex (vmPFC) to generalized, schematic memory. Here, we used a neuropsychological approach to test whether participants with focal vmPFC damage (n= 6) would show a reduced influence of schematic memory...
Find it
Spalding et al. (2018)
Ventromedial Prefrontal Cortex Is Necessary for Normal Associative Inference and Memory Integration
In
The Journal of Neuroscience, 38(15), 3767-3775
Abstract
The ability to flexibly combine existing knowledge in response to novel circumstances is highly adaptive. However, the neural correlates of flexible associative inference are not well characterized. Laboratory tests of associative inference have measured memory for overlapping pairs of studied items (e.g., AB, BC) and for nonstudied pairs with common associates (i.e., AC). Findings from functional neuroimaging and neuropsychology suggest the ventromedial prefrontal cortex (vmPFC) may be necessary for associative inference. Here, we used a neuropsychological approach to test the necessity of vmPFC for successful memory-guided associative inference in humans using an overlapping pairs...
Find it
Spencer & Raz (1995)
Differential effects of aging on memory for content and context: A meta-analysis.
In
Psychology and Aging, 10(4), 527-539
Find it
Spens & Burgess (2026)
Hippocampo-neocortical interaction as compressive retrieval-augmented generation
In
Nature Communications, 17(1)
Abstract
Abstract Many aspects of learning, memory, and problem solving involve interplay between episodic (hippocampal) and semantic (neocortical) systems, but the neural mechanisms supporting this are unclear. We present a computational model in which sequential experiences are encoded in hippocampus in compressed form and replayed to train a neocortical generative network. This network captures the gist of specific episodes and extracts statistical patterns that generalise to new situations, enabling efficient reconstruction of the past and prediction of the future. The two systems interact during encoding, recall and problem solving, with the hippocampus retrieving relevant episodic...
Find it
Spreng et al. (2009)
The Common Neural Basis of Autobiographical Memory, Prospection, Navigation, Theory of Mind, and the Default Mode: A Quantitative Meta-analysis
In
Journal of Cognitive Neuroscience, 21(3), 489-510
Abstract
AbstractA core brain network has been proposed to underlie a number of different processes, including remembering, prospection, navigation, and theory of mind [Buckner, R. L., & Carroll, D. C. Self-projection and the brain. Trends in Cognitive Sciences, 11, 49–57, 2007]. This purported network—medial prefrontal, medial-temporal, and medial and lateral parietal regions—is similar to that observed during default-mode processing and has been argued to represent self-projection [Buckner, R. L., & Carroll, D. C. Self-projection and the brain. Trends in Cognitive Sciences, 11, 49–57, 2007] or scene-construction [Hassabis, D., & Maguire, E. A. Deconstructing episodic memory with construction....
Find it
Squire & Zola (1998)
Episodic memory, semantic memory, and amnesia
In
Hippocampus, 8(3), 205-211
Find it
Squire (2004)
Memory systems of the brain: A brief history and current perspective
In
Neurobiology of Learning and Memory, 82(3), 171-177
In summary
Traces the history of the idea that memory comprises distinct systems, from nineteenth-century philosophical precursors through the mid-century experimental turn to the position, settled by about 1980, that a fundamental division separates memory available to conscious recollection from memory that is not. Describes the subsequent move beyond that dichotomy to a biologically grounded framework of multiple systems supported by the hippocampus and related structures, the amygdala, the neostriatum, and the cerebellum, and offers a current account of how these systems operate together to support behaviour.
Find it
Stadler et al. (1999)
Norms for word lists that create false memories
In
Memory & Cognition, 27(3), 494-500
Find it
Staresina et al. (2012)
Episodic Reinstatement in the Medial Temporal Lobe
In
The Journal of Neuroscience, 32(50), 18150-18156
Abstract
The essence of episodic memory is our ability to reexperience past events in great detail, even in the absence of external stimulus cues. Does the phenomenological reinstatement of past experiences go along with reinstating unique neural representations in the brain? And if so, how is this accomplished by the medial temporal lobe (MTL), a brain region intimately linked to episodic memory? Computational models suggest that such reinstatement (also termed “pattern completion”) in cortical regions is mediated by the hippocampus, a key region of the MTL. Although recent functional magnetic resonance imaging studies demonstrated reinstatement of coarse item properties like stimulus category or...
Find it
Stern & Nadel (2026)
Drawing on Memory: Dual-Trace Encoding Improves Cross-Session Recall in LLM Agents
In
arXiv
Abstract
LLM agents with persistent memory store information as flat factual records, providing little context for temporal reasoning, change tracking, or cross-session aggregation. Inspired by the drawing effect [3], we introduce dual-trace memory encoding. In this method, each stored fact is paired with a concrete scene trace, a narrative reconstruction of the moment and context in which the information was learned. The agent is forced to commit to specific contextual details during encoding, creating richer, more distinctive memory traces. Using the LongMemEval-S benchmark (4,575 sessions, 100 recall questions), we compare dual-trace encoding against a fact-only control with matched coverage...
Find it
Stramaccia et al. (2021)
Memory suppression and its deficiency in psychological disorders: A focused meta-analysis.
In
Journal of Experimental Psychology: General, 150(5), 828-850
Abstract
It is still debated whether suppressing the retrieval of unwanted memories causes forgetting and whether this constitutes a beneficial mechanism. To shed light on these 2 questions, we scrutinize the evidence for such suppression-induced forgetting (SIF) and examine whether it is deficient in psychological disorders characterized by intrusive thoughts. Specifically, we performed a focused meta-analysis of studies that have used the think/no-think procedure to test SIF in individuals either affected by psychological disorders or exhibiting high scores on related traits. Overall, across 96 effects from 25 studies, we found that avoiding retrieval leads to significant forgetting in healthy...
Find it
Sun et al. (2025)
How new data permeates LLM knowledge and how to dilute it
In
arXiv
Abstract
Large language models learn and continually learn through the accumulation of gradient-based updates, but how individual pieces of new information affect existing knowledge, leading to both beneficial generalization and problematic hallucination, remains poorly understood. We demonstrate that when learning new information, LLMs exhibit a "priming" effect: learning a new fact can cause the model to inappropriately apply that knowledge in unrelated contexts. To systematically study this phenomenon, we introduce "Outlandish," a carefully curated dataset of 1320 diverse text samples designed to probe how new knowledge permeates through an LLM's existing knowledge base. Using this dataset, we...
Find it
Svoboda et al. (2006)
The functional neuroanatomy of autobiographical memory: A meta-analysis
In
Neuropsychologia, 44(12), 2189-2208
Abstract
Autobiographical memory (AM) entails a complex set of operations, including episodic memory, self-reflection, emotion, visual imagery, attention, executive functions, and semantic processes. The heterogeneous nature of AM poses significant challenges in capturing its behavioral and neuroanatomical correlates. Investigators have recently turned their attention to the functional neuroanatomy of AM. We used the effect-location method of meta-analysis to analyze data from 24 functional imaging studies of AM. The results indicated a core neural network of left-lateralized regions, including the medial and ventrolateral prefrontal, medial and lateral temporal and retrosplenial/posterior...
Find it
Symeonidou et al. (2026)
Context reinstatement effects in younger and older adults’ memory: A meta-analysis.
In
Psychology and Aging, 41(4), 525-543
Find it
Takashiro et al. (2025)
Answer When Needed, Forget When Not: Language Models Pretend to Forget via In-Context Knowledge Unlearning
In
Findings of the Association for Computational Linguistics: ACL 2025, 24872-24885
Find it
Thompson & Spencer (1966)
Habituation: A model phenomenon for the study of neuronal substrates of behavior.
In
Psychological Review, 73(1), 16-43
Find it
Titz & Verhaeghen (2010)
Aging and directed forgetting in episodic memory: A meta-analysis.
In
Psychology and Aging, 25(2), 405-411
Find it
Todd & Marois (2004)
Capacity limit of visual short-term memory in human posterior parietal cortex
In
Nature, 428(6984), 751-754
Find it
Tompary & Davachi (2017)
Consolidation Promotes the Emergence of Representational Overlap in the Hippocampus and Medial Prefrontal Cortex
In
Neuron, 96(1), 228-241.e5
Find it
Tricomi et al. (2009)
A specific role for posterior dorsolateral striatum in human habit learning
In
European Journal of Neuroscience, 29(11), 2225-2232
Abstract
Habits are characterized by an insensitivity to their consequences and, as such, can be distinguished from goal-directed actions. The neural basis of the development of demonstrably outcome-insensitive habitual actions in humans has not been previously characterized. In this experiment, we show that extensive training on a free-operant task reduces the sensitivity of participants’ behavior to a reduction in outcome value. Analysis of functional magnetic resonance imaging data acquired during training revealed a significant increase in task-related cue sensitivity in a right posterior putamen–globus pallidus region as training progressed. These results provide evidence for a shift from...
Find it
Tsao et al. (2018)
Integrating time from experience in the lateral entorhinal cortex
In
Nature, 561(7721), 57-62
Find it
Tse et al. (2007)
Schemas and Memory Consolidation
In
Science, 316(5821), 76-82
Abstract
Reports that systems consolidation can occur very rapidly when an associative schema is already in place. Rats learned flavor-place paired associates in an event arena over several weeks; once a putative neocortical schema had developed, new associations trained on a single trial became assimilated and rapidly hippocampus-independent. Schemas also played a causal role, since the same rapid learning did not occur in an unfamiliar arena or under an inconsistent flavor-to-place mapping.
Find it
Tullis & Benjamin (2011)
On the effectiveness of self-paced learning
In
Journal of Memory and Language, 64(2), 109-118
Find it
Tulving & Thomson (1973)
Encoding specificity and retrieval processes in episodic memory.
In
Psychological Review, 80(5), 352-373
Abstract
Recent changes in pretheoretical orientation toward problems of human memory have brought with them a concern with retrieval processes, and a number of early versions of theories of retrieval have been constructed. This paper describes and evaluates explanations offered by these theories to account for the effect of extralist cuing, facilitation of recall of list items by non-list items. Experiments designed to test the currently most popular theory of retrieval, the generation-recognition theory, yielded results incompatible not only with generation-recognition models, but most other theories as well: under certain conditions subjects consistently failed to recognize many recallable list...
Find it
Tulving (1985)
How many memory systems are there?
In
American Psychologist, 40(4), 385-398
Abstract
Proposes a ternary classificatory scheme of memory in which procedural, semantic, and episodic memory constitute a monohierarchical arrangement. In this scheme, episodic memory is a specialized subsystem of semantic memory, which in turn is a specialized subsystem of procedural memory. The 3 memory systems differ from one another in a number of ways, including the kind of consciousness that characterizes their operations. The ternary scheme overlaps with dichotomies and trichotomies of memory proposed by others. Evidence for multiple systems is reviewed, and illustrative data are provided from experiments. Direct priming effects were found to be both functionally and stochastically...
Find it
Tulving & Schacter (1990)
Priming and Human Memory Systems
In
Science, 247(4940), 301-306
Abstract
Priming is a nonconscious form of human memory, which is concerned with perceptual identification of words and objects and which has only recently been recognized as separate from other forms of memory or memory systems. It is currently under intense experimental scrutiny. Evidence is converging for the proposition that priming is an expression of a perceptual representation system that operates at a pre-semantic level; it emerges early in development, and access to it lacks the kind of flexibility characteristic of other cognitive memory systems. Conceptual priming, however, seems to be based on the operations of semantic memory.
Find it
Tulving (2002)
Episodic Memory: From Mind to Brain
In
Annual Review of Psychology, 53(1), 1-25
Abstract
Episodic memory is a neurocognitive (brain/mind) system, uniquely different from other memory systems, that enables human beings to remember past experiences. The notion of episodic memory was first proposed some 30 years ago. At that time it was defined in terms of materials and tasks. It was subsequently refined and elaborated in terms of ideas such as self, subjective time, and autonoetic consciousness. This chapter provides a brief history of the concept of episodic memory, describes how it has changed (indeed greatly changed) since its inception, considers criticisms of it, and then discusses supporting evidence provided by (a) neuropsychological studies of patterns of memory...
Find it
Uttl (2008)
Transparent Meta-Analysis of Prospective Memory and Aging
In
PLoS ONE, 3(2), e1568
Find it
Vaccaro & Fleming (2018)
Thinking about thinking: A coordinate-based meta-analysis of neuroimaging studies of metacognitive judgements
In
Brain and Neuroscience Advances, 2
Abstract
Metacognition supports reflection upon and control of other cognitive processes. Despite metacognition occupying a central role in human psychology, its neural substrates remain underdetermined, partly due to study-specific differences in task domain and type of metacognitive judgement under study. It is also unclear how metacognition relates to other apparently similar abilities that depend on recursive thought such as theory of mind or mentalising. Now that neuroimaging studies of metacognition are more prevalent, we have an opportunity to characterise consistencies in neural substrates identified across different analysis types and domains. Here we used quantitative activation...
Find it
Vaidya et al. (2023)
Humans and language models diverge when predicting repeating text
In
arXiv
Abstract
Language models that are trained on the next-word prediction task have been shown to accurately model human behavior in word prediction and reading speed. In contrast with these findings, we present a scenario in which the performance of humans and LMs diverges. We collected a dataset of human next-word predictions for five stimuli that are formed by repeating spans of text. Human and GPT-2 LM predictions are strongly aligned in the first presentation of a text span, but their performance quickly diverges when memory (or in-context learning) begins to play a role. We traced the cause of this divergence to specific attention heads in a middle layer. Adding a power-law recency bias to these...
Find it
Valentin et al. (2007)
Determining the Neural Substrates of Goal-Directed Learning in the Human Brain
In
The Journal of Neuroscience, 27(15), 4019-4026
Abstract
Instrumental conditioning is considered to involve at least two distinct learning systems: a goal-directed system that learns associations between responses and the incentive value of outcomes, and a habit system that learns associations between stimuli and responses without any link to the outcome that that response engendered. Lesion studies in rodents suggest that these two distinct components of instrumental conditioning may be mediated by anatomically distinct neural systems. The aim of the present study was to determine the neural substrates of the goal-directed component of instrumental learning in humans. Nineteen human subjects were scanned with functional magnetic resonance...
Find it
van Genugten & Schacter (2024)
Automated scoring of the autobiographical interview with natural language processing
In
Behavior Research Methods, 56(3), 2243-2259
Abstract
AbstractThe autobiographical interview has been used in more than 200 studies to assess the content of autobiographical memories. In a typical experiment, participants recall memories, which are then scored manually for internal details (episodic details from the central event) and external details (largely non-episodic details). Scoring these narratives requires a significant amount of time. As a result, large studies with this procedure are often impractical, and even conducting small studies is time-consuming. To reduce scoring burden and enable larger studies, we developed an approach to automatically score responses with natural language processing. We fine-tuned an existing language...
Find it
van Kesteren et al. (2012)
How schema and novelty augment memory formation
In
Trends in Neurosciences, 35(4), 211-219
Find it
Vargha-Khadem et al. (1997)
Differential Effects of Early Hippocampal Pathology on Episodic and Semantic Memory
In
Science, 277(5324), 376-380
Abstract
Global anterograde amnesia is described in three patients with brain injuries that occurred in one case at birth, in another by age 4, and in the third at age 9. Magnetic resonance techniques revealed bilateral hippocampal pathology in all three cases. Remarkably, despite their pronounced amnesia for the episodes of everyday life, all three patients attended mainstream schools and attained levels of speech and language competence, literacy, and factual knowledge that are within the low average to average range. The findings provide support for the view that the episodic and semantic components of cognitive memory are partly dissociable, with only the episodic component being fully...
Find it
Varma et al. (2024)
A systematic review and meta-analysis of experimental methods for modulating intrusive memories following lab-analogue trauma exposure in non-clinical populations
In
Nature Human Behaviour, 8(10), 1968-1987
Find it
Vázquez-Millán et al. (2026)
The evaluation of devaluation: Deficient outcome devaluation leads to wrongly considering goal-directed actions as habits
In
Behavior Research Methods, 58(8)
Abstract
Abstract Habits are stimulus-driven responses that are produced independently of the current outcome value. They enable efficient actions in familiar contexts while freeing up cognitive resources. Habits are expected to influence behavior after substantial experience, whereas limited training leaves behavior under goal-directed control. This transition from goal-directed to habitual control has been documented in animal research but remains challenging to replicate in humans. Using a free-operant task, recent studies have suggested that human habits can be fully learned even after short training. This would produce a ceiling effect, explaining null results when the amount of training is...
Find it
Verhaeghen (2003)
Aging and vocabulary score: A meta-analysis.
In
Psychology and Aging, 18(2), 332-339
Find it
Vinken et al. (2020)
Incorporating intrinsic suppression in deep neural networks captures dynamics of adaptation in neurophysiology and perception
In
Science Advances, 6(42)
Abstract
Adding a cellular adaptation mechanism to artificial neural networks explains temporal context modulation of biological vision.
Find it
Volle et al. (2011)
The role of rostral prefrontal cortex in prospective memory: A voxel-based lesion study
In
Neuropsychologia, 49(8), 2185-2198
Find it
Wang et al. (2023)
Self-Knowledge Guided Retrieval Augmentation for Large Language Models
In
Findings of the Association for Computational Linguistics: EMNLP 2023, 10303-10315
Find it
Wang et al. (2025)
Agent Workflow Memory
In
arXiv
In summary
Language model agents struggle with long-horizon tasks involving complex action trajectories, whereas humans solve such tasks by learning reusable routines from past experience and applying them to future actions. The authors introduce Agent Workflow Memory, a method for inducing commonly reused routines and selectively supplying them to guide subsequent generation, applicable both offline from training examples and online from test queries. On two web navigation benchmarks covering more than 1000 tasks across more than 200 domains, the method improves baseline success rates by 24.6% and 51.1% relative on Mind2Web and WebArena respectively while reducing the number of steps taken, and...
Find it
Wang & Sun (2025)
Unable to Forget: Proactive Interference Reveals Working Memory Limits in LLMs Beyond Context Length
In
arXiv
Abstract
Information retrieval in Large Language Models (LLMs) is increasingly recognized as intertwined with generation capabilities rather than mere lookup. While longer contexts are often assumed to improve retrieval, the effects of intra-context interference remain understudied. To address this, we adapt the proactive interference (PI) paradigm from cognitive science, where earlier information disrupts recall of newer updates. In humans, susceptibility to such interference is inversely linked to working memory capacity. We introduce PI-LLM, an evaluation that sequentially streams semantically related key-value updates and queries only the final values. Although these final values are clearly...
Find it
Warren et al. (2014)
False Recall Is Reduced by Damage to the Ventromedial Prefrontal Cortex: Implications for Understanding the Neural Correlates of Schematic Memory
In
The Journal of Neuroscience, 34(22), 7677-7682
Abstract
Schematic memory, or contextual knowledge derived from experience (Bartlett, 1932), benefits memory function by enhancing retention and speeding learning of related information (Bransford and Johnson, 1972; Tse et al., 2007). However, schematic memory can also promote memory errors, producing false memories. One demonstration is the “false memory effect” of the Deese–Roediger–McDermott (DRM) paradigm (Roediger and McDermott, 1995): studying words that fit a common schema (e.g., cold, blizzard, winter) often produces memory for a nonstudied word (e.g., snow). We propose that frontal lobe regions that contribute to complex decision-making processes by weighting various alternatives, such as...
Find it
Wiechert et al. (2023)
Suppression-induced forgetting: a pre-registered replication of the think/no-think paradigm
In
Memory, 31(7), 989-1002
Find it
Wig et al. (2005)
Reductions in neural activity underlie behavioral components of repetition priming
In
Nature Neuroscience, 8(9), 1228-1233
Find it
Williams et al. (2007)
Autobiographical memory specificity and emotional disorder.
In
Psychological Bulletin, 133(1), 122-148
Find it
Wilson & McNaughton (1994)
Reactivation of Hippocampal Ensemble Memories During Sleep
In
Science, 265(5172), 676-679
Abstract
Simultaneous recordings were made from large ensembles of hippocampal "place cells" in three rats during spatial behavioral tasks and in slow-wave sleep preceding and following these behaviors. Cells that fired together when the animal occupied particular locations in the environment exhibited an increased tendency to fire together during subsequent sleep, in comparison to sleep episodes preceding the behavioral tasks. Cells that were inactive during behavior, or that were active but had non-overlapping spatial firing, did not show this increase. This effect, which declined gradually during each post-behavior sleep session, may result from synaptic modification during waking experience....
Find it
Wixted (2004)
The Psychology and Neuroscience of Forgetting
In
Annual Review of Psychology, 55(1), 235-269
Abstract
Traditional theories of forgetting are wedded to the notion that cue-overload interference procedures (often involving the A-B, A-C list-learning paradigm) capture the most important elements of forgetting in everyday life. However, findings from a century of work in psychology, psychopharmacology, and neuroscience converge on the notion that such procedures may pertain mainly to forgetting in the laboratory and that everyday forgetting is attributable to an altogether different form of interference. According to this idea, recently formed memories that have not yet had a chance to consolidate are vulnerable to the interfering force of mental activity and memory formation (even if the...
Find it
Wixted (2007)
Dual-process theory and signal-detection theory of recognition memory.
In
Psychological Review, 114(1), 152-176
Abstract
Two influential models of recognition memory, the unequal-variance signal-detection model and a dual-process threshold/detection model, accurately describe the receiver operating characteristic, but only the latter model can provide estimates of recollection and familiarity. Such estimates often accord with those provided by the remember-know procedure, and both methods are now widely used in the neuroscience literature to identify the brain correlates of recollection and familiarity. However, in recent years, a substantial literature has accumulated directly contrasting the signal-detection model against the threshold/detection model, and that literature is almost unanimous in its...
Find it
Wixted & Wells (2017)
The Relationship Between Eyewitness Confidence and Identification Accuracy: A New Synthesis
In
Psychological Science in the Public Interest, 18(1), 10-65
Abstract
Summary The U.S. legal system increasingly accepts the idea that the confidence expressed by an eyewitness who identified a suspect from a lineup provides little information as to the accuracy of that identification. There was a time when this pessimistic assessment was entirely reasonable because of the questionable eyewitness-identification procedures that police commonly employed. However, after more than 30 years of eyewitness-identification research, our understanding of how to properly conduct a lineup has evolved considerably, and the time seems ripe to ask how eyewitness confidence informs accuracy under more pristine testing conditions (e.g., initial, uncontaminated memory tests...
Find it
Wu et al. (2024)
LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory
In
arXiv
Abstract
Recent large language model (LLM)-driven chat assistant systems have integrated memory components to track user-assistant chat histories, enabling more accurate and personalized responses. However, their long-term memory capabilities in sustained interactions remain underexplored. We introduce LongMemEval, a comprehensive benchmark designed to evaluate five core long-term memory abilities of chat assistants: information extraction, multi-session reasoning, temporal reasoning, knowledge updates, and abstention. With 500 meticulously curated questions embedded within freely scalable user-assistant chat histories, LongMemEval presents a significant challenge to existing long-term memory...
Find it
Wu & Deng (2025)
Transitive Inference in Large Language Models and Prompting Intervention
In
ICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 1-5
Find it
Xiang et al. (2026)
Grounding Before Generalizing: How AI Differs from Humans in Causal Transfer
In
arXiv
Abstract
Extracting abstract causal structures and applying them to novel situations is a hallmark of human intelligence. While Large Language Models (LLMs) and Vision Language Models (VLMs) have shown strong performance on a wide range of reasoning tasks, their capacity for interactive causal learning -- inducing latent structures through sequential exploration and transferring them across contexts -- remains uncharacterized. Human learners accomplish such transfer after minimal exposure, whereas classical Reinforcement Learning (RL) agents fail catastrophically. Whether state-of-the-art Artificial Intelligence (AI) models possess human-like mechanisms for abstract causal structure transfer is an...
Find it
Xie et al. (2024)
Adaptive Chameleon or Stubborn Sloth: Revealing the Behavior of Large Language Models in Knowledge Conflicts
In
arXiv
Abstract
By providing external information to large language models (LLMs), tool augmentation (including retrieval augmentation) has emerged as a promising solution for addressing the limitations of LLMs' static parametric memory. However, how receptive are LLMs to such external evidence, especially when the evidence conflicts with their parametric memory? We present the first comprehensive and controlled investigation into the behavior of LLMs when encountering knowledge conflicts. We propose a systematic framework to elicit high-quality parametric memory from LLMs and construct the corresponding counter-memory, which enables us to conduct a series of controlled experiments. Our investigation...
Find it
Xiong & Wu (2025)
Do large language models learn like humans: Interleaved and spaced practice in morphological learning
In
Acta Psychologica, 260, 105518
In summary
The study compares how humans and LLMs acquire morphological patterns in an artificial language, varying blocked versus interleaved sequencing and juxtaposed versus spaced presentation across verb classification and inflection tasks. The three models tested consistently outperformed humans, showing superior few-shot learning, but diverged from human cognition in three ways. Humans benefited more from interleaving under juxtaposition, consistent with the discriminative contrast hypothesis, while model behaviour was architecture-dependent. Humans benefited from spaced exemplars while models performed better with juxtaposed input. Humans showed a metacognitive illusion, preferring blocked...
Find it
Xiong et al. (2026)
In-context superposition: human-like working memory interference in large language models
In
arXiv
Abstract
Intelligent systems must maintain and manipulate task-relevant information online to adapt to dynamic environments. This capacity, known as working memory, is fundamental to human reasoning. Yet, human working memory is strikingly limited, maintaining only three to four items in a brain with billions of neurons. Surprisingly, large language models (LLMs), despite different substrates and direct access to prior context through attention, exhibit similar working memory limitations. Why should such different systems face analogous constraints? We propose that working memory limitations reflect a general trade-off of shared representations: representational compression and reuse support...
Find it
Xu et al. (2022)
Learning to Break the Loop: Analyzing and Mitigating Repetitions for Neural Text Generation
In
arXiv
Abstract
While large-scale neural language models, such as GPT2 and BART, have achieved impressive results on various text generation tasks, they tend to get stuck in undesirable sentence-level loops with maximization-based decoding algorithms (\textit{e.g.}, greedy search). This phenomenon is counter-intuitive since there are few consecutive sentence-level repetitions in human corpora (e.g., 0.02\% in Wikitext-103). To investigate the underlying reasons for generating consecutive sentence-level repetitions, we study the relationship between the probabilities of the repetitive tokens and their previous repetitions in the context. Through our quantitative experiments, we find that 1) Language...
Find it
Xue et al. (2011)
Spaced Learning Enhances Subsequent Recognition Memory by Reducing Neural Repetition Suppression
In
Journal of Cognitive Neuroscience, 23(7), 1624-1633
Abstract
AbstractSpaced learning usually leads to better recognition memory as compared with massed learning, yet the underlying neural mechanisms remain elusive. One open question is whether the spacing effect is achieved by reducing neural repetition suppression. In this fMRI study, participants were scanned while intentionally memorizing 120 novel faces, half under the massed learning condition (i.e., four consecutive repetitions with jittered interstimulus interval) and the other half under the spaced learning condition (i.e., the four repetitions were interleaved). Recognition memory tests afterward revealed a significant spacing effect: Participants recognized more items learned under the...
Find it
Yang et al. (2025)
Do Large Language Models Perform Latent Multi-Hop Reasoning without Exploiting Shortcuts?
In
Findings of the Association for Computational Linguistics: ACL 2025, 3971-3992
Abstract
We evaluate how well Large Language Models (LLMs) latently recall and compose facts to answer multi-hop queries like “In the year Scarlett Johansson was born, the Summer Olympics were hosted in the country of”. One major challenge in such evaluation is that LLMs may have developed shortcuts by encountering the head entity “Scarlett Johansson” and the answer entity “United States” in the same training sequences or merely guess the answer based on frequency-based priors. To prevent shortcuts, we exclude test queries where the head and answer entities might have co-appeared during training. Through careful selection of relations and facts and systematic removal of cases where models might...
Find it
Yang et al. (2026)
GroupMemBench: Benchmarking LLM Agent Memory in Multi-Party Conversations
In
arXiv
Abstract
Large Language Model (LLM) agents increasingly serve as personal assistants and workplace collaborators, where their utility depends on memory systems that extract, retrieve, and apply information across long-running conversations. However, both existing memory systems and benchmarks are built around the dyadic, single-user setup, even though real deployments routinely span groups and channels with multiple users interacting with the agent and with each other. This mismatch leaves three properties of group memory unmeasured: (i) group dynamics that go beyond concatenated one-on-one chats, (ii) speaker-grounded belief tracking, where the per-user memory modeling is needed, and (iii)...
Find it
Ye & Ke (2026)
Emotion enhances relative but not absolute temporal memory: A meta-analysis
In
Psychonomic Bulletin & Review, 33(3)
Find it
Yeung (2022)
Stability and consistency of metamemory judgments within a session
In
Frontiers in Psychology, 13
Abstract
There has been much interest in assessing individual and group differences in metamemory skills. These endeavors require or would benefit from enhanced knowledge about the stability and consistency of metamemory judgments over successive trials. However, few studies have examined these aspects. Thus, the present study investigated and compared the within-session stability and consistency of three major types of metamemory judgment: judgment of learning (JOL), feeling-of-knowing judgment (FOK), and retrospective confidence judgment (RCJ), using a single-task paradigm. A total of 38 healthy young adults (17 males, 21 females) completed three trials of a face–scene associative learning task...
Find it
Yin et al. (2023)
Do Large Language Models Know What They Don’t Know?
In
Findings of the Association for Computational Linguistics: ACL 2023, 8653-8665
Abstract
Large language models (LLMs) have a wealth of knowledge that allows them to excel in various Natural Language Processing (NLP) tasks. Current research focuses on enhancing their performance within their existing knowledge. Despite their vast knowledge, LLMs are still limited by the amount of information they can accommodate and comprehend. Therefore, the ability to understand their own limitations on the unknows, referred to as self-knowledge, is of paramount importance. This study aims to evaluate LLMs’ self-knowledge by assessing their ability to identify unanswerable or unknowable questions. We introduce an automated methodology to detect uncertainty in the responses of these models,...
Find it
Yonelinas (2002)
The Nature of Recollection and Familiarity: A Review of 30 Years of Research
In
Journal of Memory and Language, 46(3), 441-517
Abstract
To account for dissociations observed in recognition memory tests, several dual-process models have been proposed that assume that recognition judgments can be based on the recollection of details about previous events or on the assessment of stimulus familiarity. In the current article, these models are examined, along with the methods that have been developed to measure recollection and familiarity. The relevant empirical literature from behavioral, neuropsychological, and neuroimaging studies is then reviewed in order to assess model predictions. Results from a variety of measurement methods, including task-dissociation and process-estimation methods, are found to lead to remarkably...
Find it
Yoon et al. (2017)
Learning and using knowledge about what other people do and don't know despite amnesia
In
Cortex, 94, 164-175
Find it
Yoon et al. (2024)
Keeping track of who knows what in multiparty conversation despite severe memory impairment
In
Neuropsychologia, 194, 108780
Find it
Yoshizawa et al. (2026)
Metacognition of ChatGPT in confidence judgements
In
Frontiers in Artificial Intelligence, 9
Abstract
Recent advances in Large Language Models (LLMs) have raised critical concerns regarding AI alignment and safety, particularly with respect to the reliability of their outputs. In humans, metacognition plays a key role in making cognition robust and adaptive. LLMs frequently express high confidence in their responses, raising the question of whether such confidence reflects human-like metacognitive capability. In this study, we systematically compared humans and GPT-4 across multiple task formats to examine how confidence relates to performance. GPT-4 consistently outperformed humans in task accuracy. This advantage was not accompanied by human-like confidence behavior: Human confidence...
Find it
Zeithamova et al. (2012)
Hippocampal and Ventral Medial Prefrontal Activation during Retrieval-Mediated Learning Supports Novel Inference
In
Neuron, 75(1), 168-179
Abstract
Memory enables flexible use of past experience to inform new behaviors. Although leading theories hypothesize that this fundamental flexibility results from the formation of integrated memory networks relating multiple experiences, the neural mechanisms that support memory integration are not well understood. Here, we demonstrate that retrieval-mediated learning, whereby prior event details are reinstated during encoding of related experiences, supports participants' ability to infer relationships between distinct events that share content. Furthermore, we show that activation changes in a functionally coupled hippocampal and ventral medial prefrontal cortical circuit track the formation...
Find it
Zeithamova et al. (2012)
The hippocampus and inferential reasoning: building memories to navigate future decisions
In
Frontiers in Human Neuroscience, 6
Abstract
A critical aspect of inferential reasoning is the ability to form relationships between items or events that were not experienced together. This review considers different perspectives on the role of the hippocampus in successful inferential reasoning during both memory encoding and retrieval. Intuitively, inference can be thought of as a logical process by which elements of individual existing memories are retrieved and recombined to answer novel questions. Such flexible retrieval is sub-served by the hippocampus and is thought to require specialized hippocampal encoding mechanisms that discretely code events such that event elements are individually accessible from memory. In addition...
Find it
Zeng et al. (2026)
LVLMs and Humans Ground Differently in Referential Communication
In
arXiv
Abstract
For generative AI agents to partner effectively with human users, the ability to accurately predict human intent is critical. But this ability to collaborate remains limited by a critical deficit: an inability to model common ground. We present a referential communication experiment with a factorial design involving director-matcher pairs (human-human, human-AI, AI-human, and AI-AI) that interact with multiple turns in repeated rounds to match pictures of objects not associated with any obvious lexicalized labels. We show that LVLMs cannot interactively generate and resolve referring expressions in a way that enables smooth communication, a crucial skill that underlies human language use....
Find it
Zhan et al. (2026)
When Memory Becomes Authority: Benchmarking Authority Collapse at the Memory Consolidation Boundary
In
arXiv
Abstract
Persistent memory allows (self-evolving) LLM agents to adapt across tasks by consolidating heterogeneous interaction histories into reusable facts, preferences, observations, and rules. Yet consolidation also imposes an implicit authorization boundary: it determines whether stored information may later be consumed as a user fact, an attested observation, or a standing instruction. We identify authority collapse, in which consolidation preserves a claim while erasing the source constraints governing its authorized use, causing the stored memory to imply greater authority than its source permits. We introduce AuthMem-Bench, a controlled paired benchmark that holds the focal claim and...
Find it
Zhang et al. (2022)
Neural correlates of transitive inference: An SDM meta-analysis on 32 fMRI studies
In
NeuroImage, 258, 119354
Find it
Zhao et al. (2024)
Set the Clock: Temporal Alignment of Pretrained Language Models
In
Findings of the Association for Computational Linguistics ACL 2024, 15015-15040
Find it
Zhao et al. (2025)
Do LLMs Recognize Your Preferences? Evaluating Personalized Preference Following in LLMs
In
arXiv
Abstract
Large Language Models (LLMs) are increasingly used as chatbots, yet their ability to personalize responses to user preferences remains limited. We introduce PrefEval, a benchmark for evaluating LLMs' ability to infer, memorize and adhere to user preferences in a long-context conversational setting. PrefEval comprises 3,000 manually curated user preference and query pairs spanning 20 topics. PrefEval contains user personalization or preference information in both explicit and implicit forms, and evaluates LLM performance using a generation and a classification task. With PrefEval, we evaluated the aforementioned preference following capabilities of 10 open-source and proprietary LLMs in...
Find it
Zhong et al. (2023)
MemoryBank: Enhancing Large Language Models with Long-Term Memory
In
Proceedings of the AAAI Conference on Artificial Intelligence, 38(17), 19724-19731
Abstract
Large Language Models (LLMs) have drastically reshaped our interactions with artificial intelligence (AI) systems, showcasing impressive performance across an extensive array of tasks. Despite this, a notable hindrance remains—the deficiency of a long-term memory mechanism within these models. This shortfall becomes increasingly evident in situations demanding sustained interaction, such as personal companion systems, psychological counseling, and secretarial assistance. Recognizing the necessity for long-term memory, we propose MemoryBank, a novel memory mechanism tailored for LLMs. MemoryBank enables the models to summon relevant memories, continually evolve through continuous memory...
Find it