Implications
There is no one-size-fits-all “perfect” AI memory
The purpose and structure of a memory system must be evaluated individually for each AI tool. What follows from that differs by who is doing the evaluating.
Implications
-
01
Good AI memory is a match between a system and a task, not a score.
There is no general answer to give. An advice relationship draws on 13 of the 29 memory mechanisms and a therapeutic one on 23, so a system that serves one of them well can still be missing most of what the other needs.
-
02
Ask what memory a job needs, not whether the memory is good.
Of the 29 mechanisms, 9 have never been compared between people and AI and 6 cannot be, so whether an AI has good memory is not a question the evidence can answer. Whether it has been tested on the ability your task rests on is.
-
03
For Clinicians: Understand Your Tools Under the Hood.
Knowing how a platform's memory is built is part of knowing the tool. What it keeps between sessions, what it quietly drops, and what it reconstructs rather than recalls all change what a patient is actually getting from it. That knowledge is teachable, too: a patient who understands what a system does with what they tell it is in a far better position to judge it. Particular attention is warranted for cross-linguistic and cross-cultural users, whose memory patterns may be poorly represented in training data.
-
04
For AI Designers: Understand Your Audience.
A use case is not a feature list. Before building for one, work out which kinds of memory the relationship actually runs on: an advice relationship draws on 13 of the 29 mechanisms and a therapeutic one on 23, and most of what therapy adds is memory processes rather than places to put things. Designing from what is technically interesting rather than from what the function requires is how a system ends up impressive at the wrong thing.
-
05
For Users: AI Literacy.
Understand, research, do the due diligence. Before trusting a platform with something that matters, find out how its memory actually works: what it retains, what it discards, and what it reconstructs rather than recalls. AI Basics explains how a model works, and the four layers a platform can hold memory in.
Cross-cutting variables
Memory Isn’t Always Clear Cut
There are a number of factors that heavily impact how memories are formed, processed, stored, and retrieved in humans. They are equally essential to consider when weighing how an AI platform handles long term memory.
Cultural Variability
In human memory, culture shapes what gets remembered, how specifically, and whether that specificity helps.
- Americans' earliest recollections arrive about six months earlier than Chinese adults' and are longer, more self-focused and more emotionally elaborate (Wang, 2001).
- Japanese viewers bind an object to its background; Americans store the two separately, though the perceptual version of the claim does not always replicate (Masuda & Nisbett, 2001; Evans et al., 2009).
- It is learned rather than innate: how caregivers reminisce predicts a child's own narrative skill two years later (Nelson & Fivush, 2004).
- Detailed recall tracks better coping among European Americans and more depressive symptoms among East Asians (Wang et al., 2018).
In AI memory, there is no cultural variable, but there is a training corpus and a filter, both of which contain cultural biases.
- GPT-2 was trained on web pages linked from Reddit, whose US users are 67% male and 64% aged 18 to 29 (Bender et al., 2021).
- Blocklist cleaning removes 42% of documents in African American English against 6.2% of White-aligned English (Dodge et al., 2021).
- Models default to the values of English-speaking and Protestant European countries (Tao et al., 2024).
- Their covert stereotypes about speakers of African American English are more negative than any human stereotypes about African Americans ever experimentally recorded, while their overt ones are positive (Hofmann et al., 2024).
Importance to discussions of memory
Human memory varies by culture, so there is no single baseline an AI system could be built to match. Its defaults are not neutral either: they come from a training corpus that over-represents some people and a filter that strips out others, so the way a model handles what it is told reflects whoever was best represented in the text it learned from.
Emotional Modulation
In human memory, how aroused you were at the time changes how strongly the event is kept.
- The amygdala modulates storage in other regions rather than running a separate emotional store (McGaugh, 2004).
- The effect is selective: arousal amplifies whatever already had priority and suppresses the rest (Mather & Sutherland, 2011).
- Acute stress, a related state, shows the same reach: it acts at encoding, after encoding, and at retrieval, where it impairs rather than helps (Shields et al., 2017).
- What it does not do is calibrate. Flashbulb memories lose accuracy at the ordinary rate while confidence in them holds (Talarico & Rubin, 2003).
In AI memory, emotional input changes the reply being written, not what is kept.
- Trauma narratives more than double GPT-4's score on a state-anxiety inventory, and mindfulness prompts only half-undo it (Ben-Zion et al., 2025).
- Emotive framing of any kind raises measured bias; neutral phrasing is the least biased (Coda-Forno et al., 2023).
- The only counterpart inside the store is an importance score from 1 to 10, assigned at write time and summed with recency and similarity (Park et al., 2023).
Importance to discussions of memory
In people, what mattered most is usually what sticks. An AI system mostly surfaces whatever is recent or resembles the current question, weighted at most by an importance score a model gave each item when it was written down, so anyone expecting it to hold on to the things that mattered should check what it actually saved.
Encoding Specificity
In human memory, what survives depends on how it was processed going in, and on whether the later question matches.
- Recognition varies between .14 and .96 on the orienting question alone, and not because the deeper task takes longer (Craik & Tulving, 1975).
- Dividing attention costs 22 to 46% at encoding but only 1 to 13% at retrieval (Craik et al., 1996).
- Rhyme encoding beats semantic encoding on a rhyme test: the match matters more than the depth (Morris et al., 1977; Tulving & Thomson, 1973).
In AI memory, that work happens at write time, and it belongs to the system built around the model rather than to the model.
- A-MEM embeds generated keywords, tags and context alongside the content; remove that step and multi-hop F1 falls from 27.02 to 9.65 (Xu et al., 2025).
- Mem0 extracts salient facts and reconciles them against the store with add, update, delete and no-op (Chhikara et al., 2025).
- HyDE manufactures the match outright, writing the document that would answer the question and searching with that instead (Gao et al., 2023).
Importance to discussions of memory
How well something is remembered depends on how it went in. For a person that is a by-product of whatever they were doing at the time; for an AI system it is a rule someone wrote in advance, which makes it worth asking what that rule keeps and what it drops.
Ben-Zion et al. (2025)
Assessing and alleviating state anxiety in large language models
In
npj Digital Medicine, 8(1)
Abstract
The use of Large Language Models (LLMs) in mental health highlights the need to understand their responses to emotional content. Previous research shows that emotion-inducing prompts can elevate “anxiety” in LLMs, affecting behavior and amplifying biases. Here, we found that traumatic narratives increased Chat-GPT-4’s reported anxiety while mindfulness-based exercises reduced it, though not to baseline. These findings suggest managing LLMs’ “emotional states” can foster safer and more ethical human-AI interactions.
Find it
Bender et al. (2021)
On the Dangers of Stochastic Parrots
In
Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, 610-623
Abstract
The past 3 years of work in NLP have been characterized by the development and deployment of ever larger language models, especially for English. BERT, its variants, GPT-2/3, and others, most recently Switch-C, have pushed the boundaries of the possible both through architectural innovations and through sheer size. Using these pretrained models and the methodology of fine-tuning them for specific tasks, researchers have extended the state of the art on a wide array of tasks as measured by leaderboards on specific benchmarks for English. In this paper, we take a step back and ask: How big is too big? What are the possible risks associated with this technology and what paths are available...
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
Coda-Forno et al. (2023)
Inducing anxiety in large language models can induce bias
In
arXiv
Abstract
Large language models (LLMs) are transforming research on machine learning while galvanizing public debates. Understanding not only when these models work well and succeed but also why they fail and misbehave is of great societal relevance. We propose to turn the lens of psychiatry, a framework used to describe and modify maladaptive behavior, to the outputs produced by these models. We focus on twelve established LLMs and subject them to a questionnaire commonly used in psychiatry. Our results show that six of the latest LLMs respond robustly to the anxiety questionnaire, producing comparable anxiety scores to humans. Moreover, the LLMs' responses can be predictably changed by using...
Find it
Craik & Tulving (1975)
Depth of processing and the retention of words in episodic memory.
In
Journal of Experimental Psychology: General, 104(3), 268-294
Abstract
Conducted 10 experiments to evaluate the notion of "depth of processing" in human memory. Undergraduate Ss were asked questions concerning the physical, phonemic, or semantic characteristics of a long series of words; this initial question phase was followed by an unexpected retention test for the words. It was hypothesized that "deeper" (semantic) questions would take longer to answer and be associated with higher retention of the target words. These ideas were confirmed by the 1st 4 experiments. Exps V-X showed (a) it is the qualitative nature of a word's encoding which determines retention, not processing time as such; and (b) retention of words given positive and negative decisions...
Find it
Craik et al. (1996)
The effects of divided attention on encoding and retrieval processes in human memory.
In
Journal of Experimental Psychology: General, 125(2), 159-180
Abstract
The authors examined the effects of divided attention (DA) at encoding and retrieval in free recall, cued recall, and recognition memory in 4 experiments. Lists of words or word pairs were presented auditorily and recalled orally; the secondary task was a visual continuous reaction-time (RT) task with manual responses. At encoding, DA was associated with large reductions in memory performance, but small increases in RT; trade-offs between memory and RT were under conscious control. In contrast, DA at retrieval resulted in small or no reductions in memory, but in comparatively larger increases in RT, especially in free recall. Memory performance was sensitive to changes in task emphasis at...
Find it
Dodge et al. (2021)
Documenting Large Webtext Corpora: A Case Study on the Colossal Clean Crawled Corpus
In
Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, 1286-1305
Abstract
Large language models have led to remarkable progress on many NLP tasks, and researchers are turning to ever-larger text corpora to train them. Some of the largest corpora available are made by scraping significant portions of the internet, and are frequently introduced with only minimal documentation. In this work we provide some of the first documentation for the Colossal Clean Crawled Corpus (C4; Raffel et al., 2020), a dataset created by applying a set of filters to a single snapshot of Common Crawl. We begin by investigating where the data came from, and find a significant amount of text from unexpected sources like patents and US military websites. Then we explore the content of the...
Find it
Evans et al. (2009)
Short Article: Scene perception and memory revealed by eye movements and receiver-operating characteristic analyses: Does a cultural difference truly exist?
In
Quarterly Journal of Experimental Psychology, 62(2), 276-285
Abstract
Cultural differences have been observed in scene perception and memory: Chinese participants purportedly attend to the background information more than did American participants. We investigated the influence of culture by recording eye movements during scene perception and while participants made recognition memory judgements. Real-world pictures with a focal object on a background were shown to both American and Chinese participants while their eye movements were recorded. Later, memory for the focal object in each scene was tested, and the relationship between the focal object (studied, new) and the background context (studied, new) was manipulated. Receiver-operating characteristic...
Find it
Gao et al. (2023)
Precise Zero-Shot Dense Retrieval without Relevance Labels
In
Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 1762-1777
Abstract
While dense retrieval has been shown to be effective and efficient across tasks and languages, it remains difficult to create effective fully zero-shot dense retrieval systems when no relevance labels are available. In this paper, we recognize the difficulty of zero-shot learning and encoding relevance. Instead, we propose to pivot through Hypothetical Document Embeddings (HyDE). Given a query, HyDE first zero-shot prompts an instruction-following language model (eg, InstructGPT) to generate a hypothetical document. The document captures relevance patterns but is “fake” and may contain hallucinations. Then, an unsupervised contrastively learned encoder (eg, Contriever) encodes the...
Find it
Hofmann et al. (2024)
AI generates covertly racist decisions about people based on their dialect
In
Nature, 633(8028), 147-154
Abstract
Hundreds of millions of people now interact with language models, with uses ranging from help with writing, to informing hiring decisions. However, these language models are known to perpetuate systematic racial prejudices, making their judgements biased in problematic ways about groups such as African Americans, , –. Although previous research has focused on overt racism in language models, social scientists have argued that racism with a more subtle character has developed over time, particularly in the United States after the civil rights movement,. It is unknown whether this covert racism manifests in language models. Here, we demonstrate that language models embody covert racism in...
Find it
Masuda & Nisbett (2001)
Attending holistically versus analytically: Comparing the context sensitivity of Japanese and Americans.
In
Journal of Personality and Social Psychology, 81(5), 922-934
Abstract
Much research indicates that East Asians, more than Americans, explain events with reference to the context. The authors examined whether East Asians also attend to the context more than Americans do. In Study 1, Japanese and Americans watched animated vignettes of underwater scenes and reported the contents. In a subsequent recognition test, they were shown previously seen objects as well as new objects, either in their original setting or in novel settings, and then were asked to judge whether they had seen the objects. Study 2 replicated the recognition task using photographs of wildlife. The results showed that the Japanese (a) made more statements about contextual information and...
Find it
Mather & Sutherland (2011)
Arousal-Biased Competition in Perception and Memory
In
Perspectives on Psychological Science, 6(2), 114-133
Abstract
Our everyday surroundings besiege us with information. The battle is for a share of our limited attention and memory, with the brain selecting the winners and discarding the losers. Previous research shows that both bottom–up and top–down factors bias competition in favor of high priority stimuli. We propose that arousal during an event increases this bias both in perception and in long-term memory of the event. Arousal-biased competition theory provides specific predictions about when arousal will enhance memory for events and when it will impair it, which accounts for some puzzling contradictions in the emotional memory literature.
Find it
McGaugh (2004)
THE AMYGDALA MODULATES THE CONSOLIDATION OF MEMORIES OF EMOTIONALLY AROUSING EXPERIENCES
In
Annual Review of Neuroscience, 27(1), 1-28
Abstract
▪ Abstract Converging findings of animal and human studies provide compelling evidence that the amygdala is critically involved in enabling us to acquire and retain lasting memories of emotional experiences. This review focuses primarily on the findings of research investigating the role of the amygdala in modulating the consolidation of long-term memories. Considerable evidence from animal studies investigating the effects of posttraining systemic or intra-amygdala infusions of hormones and drugs, as well as selective lesions of specific amygdala nuclei, indicates that (a) the amygdala mediates the memory-modulating effects of adrenal stress hormones and several classes of...
Find it
Morris et al. (1977)
Levels of processing versus transfer appropriate processing
In
Journal of Verbal Learning and Verbal Behavior, 16(5), 519-533
Abstract
Levels of processing were manipulated as a function of acquisition task and type of recognition test in three experiments. Experiment 1 showed that semantic acquisition was superior to rhyme acquisition given a standard recognition test, whereas rhyme acquisition was superior to semantic acquisition given a rhyming recognition test. The former finding supports, while the latter finding contradicts, the levels of processing claim that depth of processing leads to stronger memory traces. Experiment 2 replicated these findings using both immediate and delayed recognition tests. Experiment 3 indicated that these effects were not dependent upon the number of times a rhyme sound was presented...
Find it
Nelson & Fivush (2004)
The Emergence of Autobiographical Memory: A Social Cultural Developmental Theory.
In
Psychological Review, 111(2), 486-511
Abstract
The authors present a multicomponent dynamic developmental theory of human autobiographical memory that emerges gradually across the preschool years. The components that contribute to the process of emergence include basic memory abilities, language and narrative, adult memory talk, temporal understanding, and understanding of self and others. The authors review the empirical developmental evidence within each of these components to show how each contributes to the timing, quantity, and quality of personal memories from the early years of life. The authors then consider the relevance of the theory to explanations of childhood amnesia and how the theory accounts for and predicts the...
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
Shields et al. (2017)
The effects of acute stress on episodic memory: A meta-analysis and integrative review.
In
Psychological Bulletin, 143(6), 636-675
Abstract
A growing body of research has indicated that acute stress can critically impact memory. However, there are a number of inconsistencies in the literature, and important questions remain regarding the conditions under which stress effects emerge as well as basic questions about how stress impacts different phases of memory. In this meta-analysis, we examined 113 independent studies in humans with 6,216 participants that explored effects of stress on encoding, postencoding, retrieval, or postreactivation phases of episodic memory. The results indicated that when stress occurred prior to or during encoding it impaired memory, unless both the delay between the stressor and encoding was very...
Find it
Talarico & Rubin (2003)
Confidence, Not Consistency, Characterizes Flashbulb Memories
In
Psychological Science, 14(5), 455-461
Abstract
On September 12, 2001, 54 Duke students recorded their memory of first hearing about the terrorist attacks of September 11 and of a recent everyday event. They were tested again either 1, 6, or 32 weeks later. Consistency for the flashbulb and everyday memories did not differ, in both cases declining over time. However, ratings of vividness, recollection, and belief in the accuracy of memory declined only for everyday memories. Initial visceral emotion ratings correlated with later belief in accuracy, but not consistency, for flashbulb memories. Initial visceral emotion ratings predicted later posttraumatic stress disorder symptoms. Flashbulb memories are not special in their accuracy, as...
Find it
Tao et al. (2024)
Cultural bias and cultural alignment of large language models
In
PNAS Nexus, 3(9)
Abstract
Culture fundamentally shapes people’s reasoning, behavior, and communication. As people increasingly use generative artificial intelligence (AI) to expedite and automate personal and professional tasks, cultural values embedded in AI models may bias people’s authentic expression and contribute to the dominance of certain cultures. We conduct a disaggregated evaluation of cultural bias for five widely used large language models (OpenAI’s GPT-4o/4-turbo/4/3.5-turbo/3) by comparing the models’ responses to nationally representative survey data. All models exhibit cultural values resembling English-speaking and Protestant European countries. We test cultural prompting as a control strategy to...
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
Wang (2001)
Culture effects on adults' earliest childhood recollection and self-description: Implications for the relation between memory and the self.
In
Journal of Personality and Social Psychology, 81(2), 220-233
Abstract
American and Chinese college students (N= 256) reported their earliest childhood memory on a memory questionnaire and provided self-descriptions on a shortened 20 Statements Test (MH Kuhn & TS McPartland, 1954). The average age at earliest memory of Americans was almost 6 months earlier than that of Chinese. Americans reported lengthy, specific, self-focused, and emotionally elaborate memories; they also placed emphasis on individual attributes in describing themselves. Chinese provided brief accounts of childhood memories centering on collective activities, general routines, and emotionally neutral events; they also included a great number of social roles in their self-descriptions....
Find it
Wang et al. (2018)
Culturally Motivated Remembering: The Moderating Role of Culture for the Relation of Episodic Memory to Well-Being
In
Clinical Psychological Science, 6(6), 860-871
Abstract
Episodic memory for specific personal events is a fundamental human cognitive faculty. Yet it is variably valued across cultures and may thus have different implications for psychological well-being. In a series of studies, we investigated the consequences of cultural fit in detailed episodic recall for psychological well-being among healthy adults and children from European American and East Asian cultural backgrounds. Participants recalled recent and distant personal events (Studies 1, 2, and 3) or recent positive and negative events (Study 4). The findings showed that culture moderated the relation of episodic memory to various aspects of mental health and well-being. Whereas detailed...
Find it
Xu et al. (2025)
A-Mem: Agentic Memory for LLM Agents
In
Advances in Neural Information Processing Systems 38, 20004-20031
Abstract
While large language model (LLM) agents can effectively use external tools for complex real-world tasks, they require memory systems to leverage historical experiences. Current memory systems enable basic storage and retrieval but lack sophisticated memory organization, despite recent attempts to incorporate graph databases. Moreover, these systems' fixed operations and structures limit their adaptability across diverse tasks. To address this limitation, this paper proposes a novel agentic memory system for LLM agents that can dynamically organize memories in an agentic way. Following the basic principles of the Zettelkasten method, we designed our memory system to create interconnected...
Find it