Mechanisms
Memory Processes · Degradation
Retrieval Interference
Example
You know one thing about a new colleague, and it comes to mind instantly. You know forty things about your closest friend, and recalling which restaurant she mentioned takes a beat longer, not because the memory is weaker but because there is more attached to her.
Retrieval Interference in Humans
Definition
The cost of retrieving a particular fact rises with the number of other facts stored about the same concept. In the study that established the effect, people learned sentences pairing a person with a location, and later took longer to verify any one sentence when the person or the location appeared in three of the learned sentences rather than one (Anderson, 1974). The human literature calls this the fan effect, where fan is the number of facts known about a concept (Danker et al., 2011). The cost attaches to the concept being probed rather than to the fact itself: because the same sentences were rotated through every condition across participants, a given fact became harder to reach purely as a function of how many neighbours its concepts had accumulated (Anderson, 1974).
Scope and contested mechanism: Anderson manipulated fan only between one and three facts per concept, so nothing here extends to unbounded accumulation. The mechanism producing the cost is not settled by the reaction-time data. Anderson's own model has a serial search through the facts attached to each concept, with searches from the different concepts in a probe running in parallel, but he notes that the data are equally compatible with a parallel search whose rate slows in proportion to the number of paths, and he leaves undecided whether a search stops on a match or exhausts the set (Anderson, 1974). ACT-R gives a different account, in which a fixed pool of spreading activation is divided among a cue's associates, so each additional association weakens the link between the cue and any one fact (Danker et al., 2011).
Neural Substrate
Human fMRI locates the resolution of retrieval competition in left ventrolateral prefrontal cortex. Badre et al. (2005) varied competition four ways during semantic decisions and found that a mid-VLPFC region responded to all four, and was the only region in which a behavioural factor common to three of the manipulations accounted for variance in the signal. Danker et al. (2011) carried this into the fan paradigm itself, with a qualification. In their predefined left VLPFC region the plain effects of person fan and location fan were not reliable on their own; fan effects emerged only in interaction with which cue participants had been trained to use during learning, and were larger for the fan of the attended cue than the unattended one. What the region tracks, on this reading, is competition among the associations a person is actually using to retrieve, not the raw number of associations stored.
Within VLPFC the control functions divide. An anterior region, roughly Brodmann area 47 on the pars orbitalis of the inferior frontal gyrus, responds when a cue is too weak to bring the target to mind unaided, and its activity co-varies with lateral temporal cortex, where semantic knowledge is stored. A mid region, roughly area 45 on the pars triangularis, responds instead to conflict among representations already retrieved, and does not co-vary with temporal cortex (Badre et al., 2005; Badre & Wagner, 2007). The same split appears in lexical decision, under the labels anterior and posterior left inferior prefrontal cortex (Gold et al., 2006), and in episodic source recollection (Dobbins & Wagner, 2005). The regions these studies label are not identical. Danker et al. define theirs as a fixed box in Talairach space centred at −40, 21, 21 and described as parts of areas 45 and 46 around the inferior frontal sulcus, whereas the Badre and Wagner foci are reported in MNI space and lie ventral to that sulcus. All fall within left inferior frontal gyrus and its immediate surround, but the coordinates are not interchangeable.
AI at a Glance
In humans, the cost of retrieving a fact rises with the number of other facts attached to the concept used to reach it, and left ventrolateral prefrontal cortex tracks the resulting competition (Anderson, 1974; Badre et al., 2005; Danker et al., 2011). AI systems show the same accumulation cost at all three levels where anyone has looked for it, in the weights, in the context window, and in an external retrieval index, but nothing in them corresponds to the prefrontal control that resolves the competition, and instructions to disregard superseded material largely fail (Roberts et al., 2024; Wang & Sun, 2025; Guo et al., 2026).
- Weights: The numeric parameters fixed at the end of training, holding whatever the model absorbed from its training data (Roberts et al., 2024).
- Context window: The span of text a model can take in at once on a given request, holding both the prompt and everything said so far in a conversation (Wang & Sun, 2025).
- Retrieval index: An external store of documents searched at query time, whose results are handed to the model as text rather than held in its weights (Guo et al., 2026).
Retrieval Interference in AI
How Far It Has Been Measured
Measurement in AI: Directly Measured
AI Memory Performance: Comparable to Humans approximate provisional
Where It Lives in the System
What Has Been Measured
- 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).
Where a Mechanism Can Live
04 Product
The whole configuration, and the only thing a benchmark observes directly.
01
Weights
Across deployments
02
Context Window
Within a session
03
Memory Layer
Across sessions
Which of the three a result belongs to is inferred from it, not seen in it.
Read more
Brodmann area
Cortex in the cerebral cortex
One of the numbered cortical regions Korbinian Brodmann defined in 1909 by differences in cell layering under the microscope.
Brodmann areas are divisions of the cortex published by Korbinian Brodmann in 1909, numbered sequentially rather than by any spatial logic (the reason for the order is not settled in the historical literature), and drawn from the arrangement of cell bodies across the cortical layers in a small number of stained brains. Fifty-two were described, some of them in other primates and not identified in humans. The numbers remain the common currency for reporting locations, but the original maps were of single hemispheres, the borders vary substantially between individuals, and modern probabilistic cytoarchitectonic atlases show that a labeled coordinate is a likelihood rather than a fact.
Brodmann area 45
Cortex in the pars triangularis
The granular cytoarchitectonic area of pars triangularis, the anterior half of Broca's area.
Brodmann area 45 corresponds broadly to pars triangularis of the inferior frontal gyrus and is granular, with a well developed layer IV, while its neighbor area 44 is dysgranular and retains only a faint layer IV. In the mapping work of Amunts and colleagues its borders cross the sulci in a substantial fraction of hemispheres, so assigning a functional activation to 45 rather than 44 on sulcal grounds alone is unreliable. Functional accounts associate it with controlled semantic retrieval more than with articulation, and it is sometimes subdivided into 45a and 45p with different connections.
Inferior frontal gyrus
Cortex in the frontal lobe
The lowest lateral frontal gyrus, divided into pars opercularis, triangularis and orbitalis by the ascending rami of the lateral fissure.
The inferior frontal gyrus runs above the lateral fissure and is cut into three parts by that fissure's ascending and horizontal rami: pars opercularis behind, pars triangularis in the middle, pars orbitalis in front. In the left hemisphere the first two correspond roughly to Broca's area, and damage there is associated with non-fluent aphasia. The right IFG is repeatedly reported in response-inhibition tasks such as stop-signal, though whether its contribution is specifically inhibitory or reflects detecting behaviorally relevant signals is a live argument. Its cytoarchitectonic boundaries do not align neatly with the gross parts.
Pars orbitalis
Cortex in the inferior frontal gyrus
The front section of the inferior frontal gyrus, curving down toward the orbital surface, roughly Brodmann area 47.
Pars orbitalis is the most anterior part of the inferior frontal gyrus, lying below the anterior ramus of the Sylvian fissure and bending onto the orbital surface, where it becomes continuous with lateral orbitofrontal cortex. It is usually equated with Brodmann area 47, sometimes with 47/12 in comparative work. It is generally excluded from Broca's area, though some parcellations include it. Activity here appears in semantic tasks and in studies of emotional expression, reflecting its position between the language areas above and the orbitofrontal cortex below rather than a settled function of its own.
Pars triangularis
Cortex in the inferior frontal gyrus
The wedge-shaped middle section of the inferior frontal gyrus, cytoarchitectonically Brodmann area 45.
Pars triangularis is the triangular middle third of the inferior frontal gyrus, sitting between the anterior and ascending rami of the Sylvian fissure and corresponding to Brodmann area 45: fully granular cortex, unlike its posterior neighbor. On the left it forms the anterior half of Broca's area. It is engaged by semantic retrieval and by selection among competing meanings more than by articulation, and while functional activation for language stays left-lateralized, the anatomical asymmetry of Broca's area has not been reproducibly demonstrated. Cytoarchitectonic mapping across postmortem brains shows area 45 crossing the sulcal boundary in a substantial minority of hemispheres.
Prefrontal cortex
Cortex in the frontal lobe
The frontal cortex in front of the motor areas, defined classically by its connections with the mediodorsal thalamus rather than by any one function.
The prefrontal cortex occupies the front of the frontal lobe, anterior to premotor and motor cortex, and accounts for roughly a sixth to a quarter of human cortex depending on where the posterior border is drawn. Three competing definitions are in use (projection from the mediodorsal thalamus, the presence of a granular layer IV, and the absence of movement on electrical stimulation), and they do not enclose the same tissue, which is why rodent-to-primate homology is still argued over. It is reliably engaged by tasks with delays, rules, and conflicting options; what it contributes to those tasks is the open question, not whether it participates.
Temporal lobe
Lobe in the cerebral hemisphere
The lobe below the lateral fissure, running forward to the temporal pole and back into parietal and occipital cortex without a sharp boundary.
The temporal lobe sits beneath the lateral fissure, folded into superior, middle and inferior temporal gyri laterally, with the fusiform and parahippocampal gyri on its ventral surface and the hippocampal formation curled inside its medial wall. Heschl's gyrus, buried in the fissure's floor, carries primary auditory cortex. Its posterior boundary against parietal and occipital cortex is conventional rather than anatomical. The lobe is associated with auditory processing, object and face recognition along its ventral stream, semantic knowledge at its pole, and, through the medial structures, declarative memory.
Ventrolateral prefrontal cortex
Cortex in the prefrontal cortex
The lower lateral prefrontal cortex, overlapping the inferior frontal gyrus and Brodmann areas 44, 45, and 47.
The ventrolateral prefrontal cortex runs along the inferior part of the lateral frontal surface, below the inferior frontal sulcus, and overlaps the inferior frontal gyrus almost entirely, so in the left hemisphere it shares tissue with what is called Broca's area. It connects strongly with temporal cortex, giving it a different input profile from its dorsolateral neighbor. It is engaged during retrieval of specific memories from semantic stores, response inhibition, and selection among competing alternatives. The right-hemisphere inhibition account and the left-hemisphere selection account developed in separate literatures, and whether one control operation covers both is not settled.
Anderson (1974)
Retrieval of propositional information from long-term memory
In
Cognitive Psychology, 6(4), 451-474
Find it
Badre et al. (2005)
Dissociable Controlled Retrieval and Generalized Selection Mechanisms in Ventrolateral Prefrontal Cortex
In
Neuron, 47(6), 907-918
Find it
Badre & Wagner (2007)
Left ventrolateral prefrontal cortex and the cognitive control of memory
In
Neuropsychologia, 45(13), 2883-2901
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
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
Danker et al. (2011)
The neural correlates of competition during memory retrieval are modulated by attention to the cues
In
Neuropsychologia, 49(9), 2427-2438
Find it
Dobbins & Wagner (2005)
Domain-general and Domain-sensitive Prefrontal Mechanisms for Recollecting Events and Detecting Novelty
In
Cerebral Cortex, 15(11), 1768-1778
Find it
Gold et al. (2006)
Dissociation of Automatic and Strategic Lexical-Semantics: Functional Magnetic Resonance Imaging Evidence for Differing Roles of Multiple Frontotemporal Regions
In
Journal of Neuroscience, 26(24), 6523-6532
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
Schneider & Anderson (2012)
Modeling fan effects on the time course of associative recognition
In
Cognitive Psychology, 64(3), 127-160
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