AI Memory ≠ Human Memory
Menu

Method

How the Constructs Were Chosen and Assessed

Which memory constructs were included and why, what was left out, and the rules by which each one was assessed against current AI systems.

Part One

What Counts as a Memory Construct

Criteria for a Memory Construct

Using ⁠Tulving's (1972)⁠ episodic/semantic distinction and ⁠Squire's (2004)⁠ declarative/non-declarative taxonomy as a starting point, the memory literature on systems and processes was surveyed and each candidate was analyzed against two criteria.

  1. Criterion 1

    Established in humans

    Empirically established in humans, evidenced by

    • a stable definition with agreed-upon examples
    • an identified neural substrate or lesion dissociation
    • quantitative or meta-analytic effect sizes

  2. Criterion 2

    Implementable in AI

    Implementable and measurable in current AI memory architectures,

    • whether or not it has yet been measured

The figure below shows how candidates were sorted against the two criteria.

Included

29 Constructs

Criterion 1, established in humans: met Criterion 2, implementable in ai: met

Meet both criteria.

  • Memory Systems 8
  • Representational Structures 4
  • Memory Processes 12
  • Metacognition 5

Recast as a Condition

Cross-Cutting Factors

Criterion 1, established in humans: failed

Conditions under which constructs operate, not operations. Held to three criteria of their own.

Examples

Excluded

Excluded Constructs

Examples

  • Sensory Memory

    Criterion 1, established in humans: met Criterion 2, implementable in ai: failed

    Fails criterion 2

  • Encoding

    Criterion 1, established in humans: not the deciding test Criterion 2, implementable in ai: not the deciding test

    Wrong unit: a stage, not a function

  • Transactive Memory

    Criterion 1, established in humans: failed Criterion 2, implementable in ai: not the deciding test

    Wrong unit: a group, not a person

Humans is criterion 1, established in humans; AI is criterion 2, implementable in AI. A dash marks a candidate that was the wrong kind of thing to be put to that test.

Included Memory Constructs

Across the human memory literature, 29 constructs were identified that meet both criteria. Of these, 8 are memory systems, 4 are representational structures, 12 are memory processes (3 acquisition and transformation, 6 retrieval and surfacing, and 3 degradation), and 5 are metacognitive functions.

Constructs were grouped by the kind of purpose each one serves rather than by the declarative and non-declarative division. Memory Systems retains ⁠Squire's (2004)⁠ organization by kind of representation. Declarative and non-declarative status is kept as an attribute of each construct rather than as a division between sections.

The remaining groups distinguish operations performed on stored content (Memory Processes), organizing schemes imposed over it (Representational Structures), and second-order judgments about it (Metacognition).

Memory Systems 8

What Is Held in Memory and How It Is Stored

Immediate 1

Declarative 2

Non-Declarative 5

Representational Structures 4

Organizing Schemes for Stored Memories

Schemas 2

Person Models 2

Memory Processes 12

Operations Performed on Stored Content

Acquisition and Transformation 3

Retrieval and Surfacing 6

Degradation 3

Metacognition 5

Second-Order Judgments About Memory Itself

Attribution 2

Confidence 2

Control 1

Each tag gives the construct's status in ⁠Squire's (2004)⁠ taxonomy: declarative, non-declarative, or both. Working Memory carries none, because that taxonomy does not place it.

Cross-Cutting Factors

Some candidates were considered and discounted as constructs because they apply to more than one construct rather than standing as one of their own. They shape how several constructs operate without being functions the memory system performs. They are conditions under which memory performs, not things it performs.

The factors below are examples, not a comprehensive list. They show where the line between a construct and a factor was drawn, not every influence on memory.

A candidate was treated as a factor rather than a construct where it meets three conditions:

  1. 01 It has documented effects on two or more of the constructs.
  2. 02 It carries quantitative or meta-analytic evidence for those effects.
  3. 03 It raises a question about AI memory that the construct-level analysis cannot address.

Some factors have a well-localized neural substrate, as emotional modulation does in the basolateral amygdala (⁠McGaugh, 2004⁠); others, such as cultural variability, have none. What distinguishes a factor from a construct here is not anatomy but scope: a factor conditions how several constructs operate rather than being an operation itself.

Sample Constructs Emotion Encoding Culture
Episodic Memory Emotional Modulation affects Episodic Memory Encoding Specificity affects Episodic Memory Cultural Variability affects Episodic Memory
Autobiographical Memory Emotional Modulation affects Autobiographical Memory Cultural Variability affects Autobiographical Memory
Consolidation Emotional Modulation affects Consolidation Encoding Specificity affects Consolidation

Emotional Modulation

Emotional arousal at the time of an event modulates how strongly it is consolidated, through basolateral amygdala influence on storage in other brain regions rather than through a separate emotional memory store (⁠McGaugh, 2004⁠). The effect is selective rather than uniform, enhancing memory for central or gist information while often impairing memory for peripheral detail (⁠Kensinger, 2009⁠).

It operates at encoding, at consolidation, and at retrieval, which places it across several constructs rather than inside any one of them. ⁠Squire's (2004)⁠ own taxonomy treats amygdala modulation as acting on both declarative and non-declarative memory, consistent with its treatment here as a factor. In clinical populations the same modulation is implicated in the involuntary and fragmentary character of trauma memory (⁠Brewin, 2014⁠).

The AI Question

The closest counterpart in a current memory layer is an importance score from 1 to 10, which a model assigns when an item is written and which is summed with recency and similarity when items are retrieved (⁠Park et al., 2023⁠). It is a label the model attaches to a stored item, not a state that changes how storage happens, so what persists is still determined largely by recency, similarity, and capacity rather than by what mattered to the user.

Encoding Specificity

What is retained depends on how material was processed at input, not merely on whether it was attended. Semantic or elaborative processing produces substantially better retention than perceptual processing of the same material, an effect established by the levels-of-processing framework (⁠Craik & Lockhart, 1972⁠) and confirmed by decades of subsequent work. The advantage is not absolute, because retention also depends on how well the operations at retrieval match those at encoding: rhyme encoding outperforms semantic encoding on a rhyme test (⁠Morris et al., 1977⁠; ⁠Tulving & Thomson, 1973⁠). Selection is further constrained upstream by attentional capacity, since material outside the focus of attention is largely not encoded at all.

The factor determines what the constructs it touches have available to operate on, and which cues will later reach it.

The AI Question

The write path in a memory layer is a policy decision rather than a capacity constraint. A system can in principle retain everything, so what it retains reflects a design choice about relevance, made once, by someone other than the user.

Cultural Variability

Wang's research program documents systematic cross-cultural differences in the content, structure, valence, accessibility, developmental origin, and functional use of autobiographical memory (⁠Wang, 2016⁠), operating through self-goals, language, emotion knowledge, and perceptual style (⁠Wang, 2021⁠). The differences are measurable: Americans' earliest recollections date from about six months earlier than Chinese adults' and are longer, more self-focused, and more emotionally elaborate (⁠Wang, 2001⁠), and the same detailed episodic recall tracks better coping among European Americans but more depressive symptoms among East Asians (⁠Wang et al., 2018⁠).

The AI Question

A current AI system has no cultural variable, but it has a training corpus and a filter, and both carry cultural bias. 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 then removes 42% of documents in African American English against 6.2% of White-aligned English (⁠Dodge et al., 2021⁠); and models default to the values of English-speaking and Protestant European countries (⁠Tao et al., 2024⁠). Human memory has no single cultural baseline, yet cultural factors heavily impact memory and reasoning. Having AI that defaults heavily to certain cultures is a critical issue that merits careful research and consideration.

Samples of Excluded Memory Constructs

Below are examples of candidates from the memory literature which were considered and excluded, either for failing one of the two criteria or for being the wrong unit of comparison. They are listed because their exclusion reflects the boundaries of the criteria rather than the limits of the review.

Sensory Memory

Fails criterion 2

Sensory memory meets the first criterion, with an established paradigm (⁠Sperling, 1960⁠), documented decay parameters, and reported effect sizes, but fails the second. Its counterpart in an AI system is input tokenization, which is not memory in any sense examined here. The exclusion is noted explicitly because existing surveys of AI memory routinely include sensory memory as a fifth system, on the strength of that mapping.

Encoding

Wrong unit: a stage, not a function

Encoding was excluded as a processing stage rather than a function. Stages describe when an operation occurs rather than what it accomplishes, which makes them the wrong unit for comparison across substrates. Encoding specificity is treated instead as a cross-cutting factor.

Transactive Memory

Wrong unit: a group, not a person

Transactive memory was excluded as a different unit of analysis rather than a different construct. Wegner's account describes encoding, storage, and retrieval distributed across members of a group, with individuals holding a directory of who knows what (⁠Wegner et al., 1991⁠). Because the system is the group rather than the person, it carries no neural substrate of the kind the first criterion requires, and it is not a factor modulating individual memory processes in the way the cross-cutting factors are.

It is retained as a framing device for the use-case comparison, where the relevant question is not what memory functions an AI system should have but what a transactive partner requires in order to be a good one.

Part Two

How AI Memory Was Assessed

Product and Layer

The unit a user actually encounters is a product: a deployed assistant such as Claude, ChatGPT, or Gemini. A product is a configuration rather than a component. It comprises a particular model, a particular set of memory rules, a system prompt, tool access, sampling parameters, and a set of defaults. An "out-of-the-box" AI refers to a product under those default settings, which is the configuration in which most users meet a general-purpose assistant.

Within a product, memory is distributed across three separately configurable layers, distinguished by how long what they hold persists. Because they behave differently and are studied unevenly, each claim on this site is tagged to the layer it concerns, or to the product where no single layer applies. Components that shape memory behavior without persisting on a horizon of their own, including the system prompt and sampling parameters, are treated as configuration and discussed at the product level.

A Published Benchmark observes this

04 Product

A configuration: a model, memory rules, a system prompt, tools, sampling parameters, and defaults.

01

Weights

Across deployments

processes

02

Context Window

Within a session

writes from, reads back into

03

Memory Layer

Across sessions

Inferred

Which of the three a failure belongs to. Drawn from the result, not observed in it.

01 Weights Across deployments

What training put into the model's parameters. Fixed during inference, so nothing that happens within a conversation changes them, but modifiable between deployments through further training.

Example

Pretraining and post-training, including RLHF; fine-tuning a base model on domain data.

02 Context Window Within a session

Everything visible to the model during a single response: the conversation so far, the system prompt, and any documents supplied. Context is bounded, and when a session exceeds that bound the excess is truncated, dropped from the front of a sliding window, or compressed into a summary that replaces the original text. Whatever remains is discarded entirely when the session ends.

This within-session compression is the closest functional analog to consolidation available in these systems, and like consolidation it is lossy, rule-governed, and determinative of what survives.

Example

The context window; in-context learning; summarization or compaction of an overlong conversation.

03 Memory Layer Across sessions

An external store that persists between sessions, together with the rules governing what is written to it and what is retrieved back into context.

Example

Retrieval-augmented generation over a vector database; agentic memory frameworks such as Mem0; a user-facing "remember this" feature.

Two Kinds of Memory Layer

Memory layer implementations divide along a line that matters for the comparisons on this site. They are functional analogs of episodic and semantic organization respectively, and they are not interchangeable.

Property Raw session transcripts Facts extracted at write time
Keeps The occasion on which something was said A distilled statement
Discards Nothing at write time The occasion
Fails by Retrieving the wrong occasion Having flattened it before retrieval was ever attempted
Functional analog Episodic organization Semantic organization

What a Benchmark Measures

The layers are not independent. The weights process whatever is in the context, and the memory layer both writes from context and reads back into it, so retrieved content is subject to the same bound and the same compression as anything else. A claim is product level when it concerns the assembled system rather than a single layer, whether because it describes such an interaction or because it describes which configuration was selected. Standing Constraints are a recurring case: prohibitions and preferences are enforced through the system prompt and expressed subject to sampling parameters, which makes them product-level claims throughout.

This carries evidentiary weight. Published benchmarks evaluate products, not layers. When one reports a memory failure, what was measured is a configuration, and attributing that failure to the memory layer rather than to the weights or to context management is an inference drawn from the result rather than something the result observed.

For a plain-language account of the layers, see AI Basics.

Level of Measurement

Each construct was assessed for how far it has been studied in existing AI systems, and placed in one of three states. An ability tested in some of the places an AI system holds information and not others is counted as Indirectly Measured.

Directly Measured

14 of 29 constructs

Definition
A published study tests AI on this ability by name, or gives AI the same experiment psychologists use to test it in people.
Assignment requires
Citing that study and the human experiment it follows.
Example
ImplicitMemBench on procedural memory, priming, and conditioning.

Indirectly Measured

15 of 29 constructs

Definition
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.
Assignment requires
Naming the AI measure and the human ability it corresponds to. An ability measured in some places and not others is placed here, since a partial result is what that is.
Example
Semantic entropy (⁠Farquhar et al., 2024⁠) operationalizes the same resampling-and-dispersion logic as ⁠Koriat's (1993)⁠ self-consistency account, with no cross-citation between the literatures.

AI Memory Performance

Each construct also received an AI memory performance rating: a coarse judgment of how the AI result compares with the human one. The ratings share their structure with the levels of measurement, and the states that are not results sit last in both. A rating can rest on a direct comparison, where people and AI took the same test, or on an approximate one, where different tests of the same ability can be put on a common scale; both are set out in the scoring criteria, along with the findings behind every rating. The scale runs in both directions from Comparable, and its last two steps are not results at all: Not Yet Compared, where a comparison could be run but has not been, and Comparison Not Possible, where no fair test could put both sides on one scale.

Exceeds Human

1 of 29 constructs

Definition
AI did better than the normal range for people.
Assignment requires
A direct or approximate comparison, and the size of the margin. A result above the human range on one structure or condition, among others within it, is Comparable, as with Priming.
Example
Semantic Knowledge. On the general-knowledge subtest of an adult intelligence scale, scored against its published norms by clinical psychologists, nine of ten systems reached the top scaled score, three standard deviations above the average adult (⁠Galatzer-Levy et al., 2024⁠).

Comparable to Humans

5 of 29 constructs

Definition
AI fell within the normal range for people.
Assignment requires
A direct or approximate comparison, the human range, and where the AI result sits in it.
Example
Priming. Given the materials of a 1998 sentence-completion experiment, ChatGPT reused the primed structure with an effect of about 9 points with a different verb and 14 with the same one, inside the human means of about 5 and 18; Vicuna's effects, about 21 and 38, sit above them (⁠Cai et al., 2024⁠).

Below Human

8 of 29 constructs

Definition
AI did worse than the normal range for people.
Assignment requires
A direct or approximate comparison, the size of the gap, and the limits of the human results.
Example
Classical Conditioning. On the same 100 items, scored the same way, the best model reached 69.7 percent against 100 percent for all five people tested, a baseline small enough and easy enough to be worth stating (⁠Qin et al., 2026⁠).

Not Yet Compared

9 of 29 constructs

not on the scale

Definition
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.
Assignment requires
Naming the test that could be run on both sides, and what is missing now: a human group, an AI result, or a shared scale that a study could supply.
Example
Integrative Inference. A long-conversation benchmark asks questions that can only be answered by combining facts from different sessions, and prints a human score of 85.8 F1 beside model scores of 12.8 to 42.0. It never says how many people answered, or whether they could reread the conversations, so there is no documented human result to compare against (⁠Maharana et al., 2024⁠).

Comparison Not Possible

6 of 29 constructs

not on the scale

Definition
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.
Assignment requires
Naming why no shared test could be built, not only that none has been.
Example
Autobiographical Memory. The ability is memory for one's own life, organized around the self; an AI system has no life of its own to remember, so the tests that exist give a model a character's past to answer about instead, which no human test corresponds to.

When Results Differ

A construct is often assessed at more than one layer, or on more than one paradigm, and the results do not always agree. Each result is judged separately, and the construct carries a single rating collapsed from them by three rules. The full breakdown remains on each construct's own page.

  1. 01 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. 02 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. 03 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.

This is the performance counterpart of the measurement rule above, where a construct reporting different states at different layers is counted as Indirectly Measured.

Episodic Memory

  • Passage order judged from memory, after training on the books Below Human
  • The same passages judged with the book still in view Comparable to Humans
AI Memory Performance Below Human

Recognition

  • Familiarity: studied sentences found less surprising than new ones Exceeds Human
  • Recollection: remember/know or retrieving the study context Not Yet Compared
AI Memory Performance Not Yet Compared

Instrumental Learning

  • Learning from rewards, with the history in view Comparable to Humans
  • Habit and outcome devaluation Not Yet Compared
AI Memory Performance Comparable to Humans

Who Coded the Constructs

Every construct's inclusion, level of measurement, and AI memory performance rating, and every Table 2 code, was assigned by the first author. They have not yet been coded independently by a second rater, so no inter-rater reliability is reported. Each assignment is argued in full on the construct's own page, with its sources, so that it can be checked, and ratings that have been re-coded or given ahead of their write-up are marked provisional.

A Note on a Young Field

The psychology of memory is an established field with a lengthy history of empirical evaluation; research on AI memory is younger and less developed. The AI half of this work represents the most relevant research available at the time of writing, and makes no claim to be a comprehensive overview of a rapidly evolving field.

The asymmetry in evidentiary maturity between the human and AI halves of each entry is itself part of what this work documents: not only that the AI literature is younger, but that it measures at a coarser grain.