RESEARCH DOSSIER
Artificial Episodic Memory Architectures
Abstract. Artificial episodic memory architectures preserve, organize, retrieve, and apply records of particular past events in intelligent artificial systems. This fundamental research monograph provides a critical narrative synthesis of human episodic-memory research, comparative episodic-like behavior, differentiable memory systems, neural episodic control, and modern language-agent architectures. It defines operational criteria for event specificity, temporal order, source attribution, retention, and memory-guided behavior, distinguishing observable machine competence from autonoetic experience in human recollection. Original experiments, architecture studies, and benchmark evidence are examined across Neural Turing Machines, memory networks, Neural Episodic Control, MemGPT, Generative Agents, Reflexion, ReadAgent, HippoRAG, MemoryBank, LoCoMo, LongMemEval, MemoryAgentBench, and LongMemEval-V2. The analysis tracks how segmentation, indexing, source links, summarization, memory budgets, and retrieval policies affect the fidelity and practical usefulness of episode records. It identifies distinct error sources in storage, retrieval, chronological reasoning, evidence attribution, and generative reconstruction. Experimental limitations include controlled synthetic histories, benchmark contamination, dependence on model and retrieval configurations, weak longitudinal generalization, and the lack of independent evidence concerning artificial subjective re-experiencing. The synthesis develops falsifiable research protocols for source-sensitive recall, temporal corrections, counterfactual histories, selective forgetting, privacy controls, and transfer across model versions. Conclusions concern the measured functions and limitations of artificial memory architectures; the manuscript reports no original experiment, no meta-analysis, and no registered systematic-review protocol.
Keywords: artificial episodic memory; episodic-like memory; artificial cognition; event encoding; temporal reasoning; source monitoring; retrieval-augmented generation; agent memory; long-term memory; memory consolidation; episodic control; source provenance; memory interference; LoCoMo; LongMemEval; memory benchmarks; generative agents; Artificial Era
The Scientific Object and Its Boundaries
Artificial episodic memory architectures are engineered systems that encode, organize, retrieve, and use records of particular past interactions or events, including their content, temporal location, contextual relations, and evidential source. The research object is the architecture's observable capacity to preserve and reconstruct episode-specific information across intervening operations and sessions. An episode might be a user instruction at a known time, an action in a simulated environment, a tool result, or a sequence of observations and consequences. Unlike general-purpose knowledge storage, an episode record must permit discriminating one occurrence from another and answering questions about what happened, in what order, and on whose evidence. Architecture, storage, retrieval policy, and task behavior are separable variables. A model may generate a plausible narrative about a past episode without recovering a trace of it; a database may preserve a perfect trace that an agent never retrieves when needed. Both failures matter for a psychological-functional account of artificial memory Tulving (2002) Park et al. (2023) Wu et al. (2025).
The scope of this dossier is event-specific encoding, segmentation, indexing, retrieval, temporal ordering, source provenance, interference, error correction, and memory-guided action in artificial agents. Parametric model weights, context windows, vector databases, and operating-system style memory tiers enter the analysis as possible substrates, but their general engineering comparison belongs principally to the neighboring Artificial Memory Systems and Storage Architectures dossier. Likewise, long-term personality, identity, continual parameter learning, and human attachment require different dependent variables. This monograph evaluates the narrow but substantial question of whether an artificial system can maintain an accountable history of distinct events and use that history appropriately after time and context have changed Packer et al. (2023) Zhang et al. (2025).
The method is a critical narrative synthesis of foundational cognitive psychology, comparative behavior, computational memory research, and published empirical evaluations of language agents. The bibliography combines primary experiments with theoretical accounts and reproducible benchmark reports. No experiments were conducted for this manuscript, and the evidence was not assembled through a preregistered systematic-review protocol. Throughout, statements about functioning software refer to measured computational or behavioral performance; statements about human recollection and its subjective features retain the definitions and evidence standards of their original discipline Clayton & Dickinson (1998) Tulving (2002) Hu et al. (2026).
Episodic Memory in Human Cognitive Psychology
Endel Tulving's seminal 1972 distinction between episodic and semantic memory became a sustained research program refined in Elements of Episodic Memory (1983) and later reviews. Semantic memory concerns knowledge that can be accessed without reinstating a particular acquisition event; episodic recollection concerns a particular occurrence embedded in a temporal and contextual frame. Tulving's mature account incorporates autonoetic consciousness, the subjective awareness of remembering one's own past. The evolution of the concept matters because computer-science authors often borrow the earlier event-content sense while omitting its later phenomenological requirements. The two usages need explicit operational translation rather than informal equivalence Tulving (1983) Tulving (2002).
In cognitive neuroscience, hippocampal damage has provided evidence that relations among people, places, and episodes can be disrupted even when other forms of knowledge remain available. Hassabis and colleagues found difficulties in constructing coherent imagined scenes in patients with hippocampal amnesia; their work connects episodic reconstruction to scene construction while leaving important interpretive questions about the organization of impaired functions. Spens and Burgess later proposed a generative computational model in which memory replay and schema-based expectations jointly shape reconstructions. These models show why remembering has both a trace-dependent and an inferential component. They provide conceptual tools for machine architectures, but do not establish that software memories engage biological processes or phenomenological experience Hassabis et al. (2007) Spens & Burgess (2024).
Human episodic memory is consequently an important scientific source of hypotheses about binding, retrieval cues, order, and distortion. It does not furnish a ready-made pass/fail certification of artificial subjectivity. A useful analysis distinguishes content criterion, referring to integrated what-where-when information; structural criterion, referring to linked event-specific representations; behavioral criterion, referring to flexible downstream use; and experiential criterion, referring to subjective re-experiencing. Artificial agents can be evaluated directly on the first three with suitable experiments. The fourth requires a separate theory and independently validated evidence Tulving (2002) Clayton & Dickinson (1998).
Episodic-Like Behavior and Comparative Criteria
Comparative psychology established a precedent for studying event memory when first-person testimony is unavailable. Clayton and Dickinson's 1998 scrub-jay experiment combined a food-caching ecology with manipulations of location, food type, and elapsed time. The birds' recovery choices varied with the perishability of the stored food and the retention interval. The result provided experimental evidence for memory for what was cached, where it was placed, and when the caching occurred. The researchers' term episodic-like memory signaled that behavioral evidence did not settle whether the animals experienced subjective mental time travel Clayton & Dickinson (1998).
This distinction is directly transferable to artificial agents as a research design principle. An agent can be exposed to unique episodes that differ along factorial combinations of content, source, place, order, and delay. Later tests can require a specific remembered episode, including counterfactual lures built from familiar elements. The central measurement is not merely whether an output sounds autobiographical but whether answers track the manipulated event structure more reliably than semantic priors, recent context, or generalized patterns. A synthetic system with a perfect event log might succeed because retrieval is deterministic; a generative agent might fail when its memory summary merges two similar occasions. The mechanisms differ, even if both produce fluent first-person reports Clayton & Dickinson (1998) Maharana et al. (2024).
Comparative reasoning also cautions against equating task success with an identical implementation. Scrub-jay cache recovery is a biologically situated behavior; artificial conversation recall is a computationally mediated task. Their comparison is useful at the level of observables and experimental control. It can reveal whether the tested function depends on content, context, and delay, without assuming shared anatomy, learning history, motivational structure, or subjective point of view. Construct validity improves when the exact sense of episodic in an evaluation is declared before results are interpreted Tulving (2002) Hu et al. (2026).
Historical Development of Machine Memory
The trajectory toward contemporary artificial episodic memory joins several historically distinct lines. Traditional software stored records externally and accessed them through exact identifiers. Connectionist models explored associative retrieval and learned representations. Neural Turing Machines, introduced by Graves, Wayne, and Danihelka in 2014, coupled a trainable controller to an addressable external memory. End-to-end Memory Networks subsequently developed learned attention over stored representations, while the differentiable neural computer demonstrated neural control of structured memory operations. These systems expanded the mechanisms by which an agent could read from and write to a memory store; they did not automatically create temporally grounded autobiographical event histories Graves et al. (2014) Sukhbaatar et al. (2015) Graves et al. (2016).
A further line explicitly used remembered experiences to guide action. Pritzel and colleagues' Neural Episodic Control employed a buffer of prior state representations and rapidly updated value estimates, enabling fast learning in reinforcement-learning environments. The word episodic here denoted access to past states and outcomes useful for decisions, not a verbal recollection of a particular lived event. More recent language-based systems integrate natural-language observation streams, retrieval, reflections, or self-management of context. The scientific progression is best described as a widening range of memory functions and interfaces, because no single architectural family is a universal successor to the others Pritzel et al. (2017) Park et al. (2023) Packer et al. (2023).
The transition to long-running language agents made the distinction between a transient prompt and persistent episode records practically urgent. A long context can carry past events until truncated, but sustained interactions require selective preservation and later readmission of relevant information. MemGPT proposed operating-system-inspired memory management between tiers; Generative Agents used a memory stream and higher-level reflections to support behavior in a simulated town. These designs created opportunities to evaluate history-sensitive conduct, provenance, and temporal coherence across multiple actions. Their published results demonstrate particular implemented patterns under their evaluation conditions, rather than a completed general theory of artificial remembering Packer et al. (2023) Park et al. (2023).
Operational Definitions and Levels of Explanation
Four descriptive levels must be kept apart. At the data level, a trace is a durable entry or reconstructable state associated with an occurrence. At the representational level, the trace binds fields such as actor, object, action, time, and source. At the computational level, an encoder, store, index, and retriever transform and select those traces. At the behavioral level, the agent uses retrieved evidence to answer, plan, revise a belief, or avoid repeating an error. Successful retrieval at one level does not guarantee accuracy at another: a correct raw record can be omitted by ranking; a correct retrieval can be distorted during answer synthesis; and a correct answer can arise from a shortcut unrelated to the episode. Experimental analysis needs instrumentation at all four levels Graves et al. (2016) Wu et al. (2025).
Episode specificity can be operationalized as an ability to differentiate two occurrences with overlapping content. Temporal organization means recovering chronology, intervals, and changes in the status of statements. Source attribution means identifying whether the event came from a user, external tool, internal computation, simulated character, or a subsequently corrected report. Retention means that an earlier episode remains accessible after unrelated interactions or explicit memory-management operations. Flexible use means correct application of episode evidence to a new question or action. These definitions yield distinct measurable outcomes and resist collapsing all memory evaluation into one aggregate response-accuracy score Johnson et al. (1993) Wu et al. (2025) Wu et al. (2026).
An important distinction is between episodic storage and episodic competence. A system may retain full transcripts yet fail source discrimination; it may store compressed experiences yet answer many questions correctly. The explanatory claim must specify which mechanism produced which observed behavior. Architectural ablations—removing timestamps, retrieval, summarization, or evidence verification while preserving the rest of the system—are particularly informative because they address causal contribution more directly than comparisons among unrelated models. Even then, the generality of an ablation result depends on the task distribution, baseline strength, and independence of evaluation data Park et al. (2023) Maharana et al. (2024).
Research Questions, Evidence Hierarchy, and Review Method
The monograph is organized around interlocking questions: What qualifies as one artificial episode? Which representations preserve the source and temporal structure of that event? How do storage and retrieval choices affect accuracy under interference? What makes recalled evidence useful for decisions and interactions? Which benchmarks distinguish recollection from inference, contamination, and unsupported generation? And which research claims remain unresolved when present systems are tested over weeks, versions, tasks, and changing environments? The evidence review takes these questions as organizing variables rather than applying a single human-memory taxonomy to every machine implementation Zhang et al. (2025) Hu et al. (2026).
Primary scientific studies have priority when attributing an experimental design or reported finding. Human memory studies inform concepts and validity threats; computational papers establish particular mechanisms; benchmark papers establish the limits of the tested tasks. Peer-reviewed proceedings and journal articles are treated differently from preliminary arXiv manuscripts. For a preprint, findings are attributed to its authors as reported results and remain subject to subsequent publication and independent replication. Source metadata and bibliographic links are retained to make the synthesis inspectable. Neither a benchmark developer's favorable score nor a model's fluent autobiographical language is treated as independent evidence of conscious recollection Tulving (2002) Wu et al. (2025) Wu et al. (2026).
The evidence hierarchy favors controlled manipulation and independent replication over single anecdotal demonstrations. Reported effects need attention to baseline performance, dataset composition, leakage, confidence intervals, task validity, and external transfer. Naturalistic long-history tests add ecological realism but can conceal ground-truth ambiguities; small synthetic episodes permit precision yet may overestimate real-world generalization. The highest-value studies pair explicit event logs with delayed retrieval and downstream decisions, then examine failures through trace-level auditing. This framework guides the subsequent critical synthesis and confines conclusions to the tested features of artificial systems Maharana et al. (2024) Liu et al. (2024) Wu et al. (2025).
Event Segmentation and the Formation of Episodes
Any episodic architecture must decide where one event ends and another begins. In a conversational agent, an event may be a single message, a tool call, a task attempt, a completed workflow, or a socially meaningful exchange spanning several turns. Granularity affects what can be retrieved later: message-level storage preserves precise wording but produces numerous fragments; session-level summaries preserve a storyline while concealing specific changes, corrections, and sources. In a foundational event-cognition model, Zacks and colleagues linked the perception of event boundaries to prediction and updating of mental event representations. Their account motivates segment-level hypotheses for machine traces, while the engineering choice must still be tested against artificial tasks with explicit ground truth. No universal segmentation length follows from cognitive analogy Zacks et al. (2007) Spens & Burgess (2024) Lee et al. (2024) Wu et al. (2025).
An event encoder can store raw observation, normalized participants, action type, input and output identifiers, physical or simulated location, time of observation, and confidence in the observation. A linked causal chain can describe whether one action preceded, enabled, or responded to another. Such structure supports questions that natural-language similarity alone may miss, including which tool returned a result before a user revised a requirement. The stored text must also identify the boundaries between a direct observation, a system inference, and an external report. When these categories are fused into one paragraph, the agent can later treat a conjecture as a witnessed occurrence Johnson et al. (1993) Park et al. (2023).
Selective encoding creates a predictable tradeoff between cost and future utility. Salience-based policies may prefer unusual, recent, goal-relevant, or emotionally marked information, as in various agent-inspired memory mechanisms. Yet the importance of an episode is often revealed only by a future question. An apparently routine instruction can become decisive weeks later when an exception is disputed. Therefore a research comparison should report storage budgets, retained raw evidence, deletion rates, and how often important details survive compaction. Policies can be compared under equal memory budgets rather than different and unreported amounts of retained text Park et al. (2023) Zhong et al. (2024) Lee et al. (2024).
Temporal Representation, Ordering, and Change
Artificial event memory requires a model of time sufficiently expressive for the questions the agent must answer. A timestamp can represent when a record was written, when a reported event occurred, or when the system learned about it; these are often different values. In a delayed conversation, the user may say that a meeting happened last Tuesday, creating an event time in the past and an observation time in the present. An architecture that stores only the ingestion timestamp cannot reliably reconstruct the historical order. Multiple clocks, normalized time zones, and explicit uncertainty intervals support more faithful temporal reasoning Maharana et al. (2024) Wu et al. (2025).
Chronology is not equivalent to causality. If a correction follows an original statement, the later statement may supersede it for current decisions, while the earlier claim remains essential for historical explanation. Memory must represent this relation instead of overwriting the original without trace. Graph structures can encode temporal succession and revision edges; append-only logs can preserve record provenance while derived current-state views update separately. Experiments should manipulate order, contradictory updates, event backdating, and questions requiring both the old belief and the presently valid one. A model that answers the most recent fact correctly may still fail a question asking what was believed before the correction Wu et al. (2025) Wu et al. (2026).
Temporal retrieval must also handle vague expressions such as before that, during the first attempt, or after the second failure. These questions demand event linking, not keyword matching. LoCoMo embeds long conversations with temporal event graphs, and LongMemEval specifically includes temporal reasoning and knowledge-update demands. Their use is informative because ordinary fact retrieval can be correct while event ordering remains wrong. Evaluation should further distinguish absolute timestamp questions, relative-order questions, interval calculations, and temporal counterfactuals, reporting accuracy separately for each class Maharana et al. (2024) Wu et al. (2025).
Source Attribution and Evidence Provenance
Remembering that a proposition appeared in a history is different from remembering who asserted it and what evidence supported it. Johnson, Hashtroudi, and Lindsay's source-monitoring framework documents that human recollection involves judgments about origins of remembered information. For artificial systems the analogous engineering problem is to label data origin, speaker identity, authority, transmission route, and transformation history. A statement from a user, a retrieved web page, a generated hypothesis, and a tool's confirmed return may be textually similar; their evidential roles are different. Correct source attribution is consequently an independent dimension of episodic performance Johnson et al. (1993).
A robust episode record can include source identifiers, document versions, tool execution IDs, observation timestamps, quotation spans, and an immutable link to original material. Transformations such as summarization should preserve a traceable relation to the source record. This design supports source-sensitive answers including I observed, the user reported, the external service returned, and the system inferred. It also permits corrections when a document is withdrawn or a tool result is superseded. The system's internal claim of certainty should be checked against source type and provenance rather than treated as an intrinsic measure of truth Wu et al. (2025) Johnson et al. (1993).
Controlled source-swap tests expose failures that ordinary question-answering benchmarks overlook. An evaluator can present two agents with the same propositional content but different speaker assignments, then ask who supplied a claim or what an earlier assistant merely suggested. Another test may insert an unreliable document that repeats a true statement while adding false metadata. A faithful architecture should preserve the distinction between retrieved content and verified record. Such protocols are essential for social and institutional uses, because an inaccurate attribution can misassign responsibility even when the extracted factual content is correct Maharana et al. (2024) Wu et al. (2026).
Representational Binding and Multimodal Episodes
An episode may integrate text, images, environmental state, tool results, and agent actions. Multimodal encoding must bind features to the same event without mistakenly joining similar observations from different occasions. Natural-language descriptions are easy to inspect but may omit visual details; embeddings support approximate similarity but generally do not provide explicit causal or temporal guarantees; structured records preserve identifiers but can be brittle when natural scenes are ambiguous. Combining these forms can be useful, provided tests identify which representation supplies the decisive evidence Graves et al. (2016) Maharana et al. (2024).
Human work on event reconstruction underscores the importance of relational structure. The experimental finding that hippocampal amnesia can impair imagined scene construction and the computational proposal that memory combines unique features with prior schemas motivate questions about how artificial systems bind relations and fill gaps. Yet neural models of biological episodic reconstruction are theoretical accounts of cognition, not validated blueprints for language-agent databases. In an artificial setting the critical evidence is whether a unique combination of features survives compression and can be recovered without importing details from an unrelated but similar event Hassabis et al. (2007) Spens & Burgess (2024).
LoCoMo's design includes image sharing and reactions within extended persona-driven conversations. It therefore extends memory testing beyond plain lists of text facts, though it remains a constructed benchmark with human curation and artificial conversational participants. A stronger multimodal test would store two visually similar artifacts mentioned in different sessions, alter one, and ask the system to attribute the changed feature to the correct encounter. Ground-truth image identifiers and update histories would permit separating perceptual error from storage, retrieval, and narration error Maharana et al. (2024).
Storage Layers, Persistence, and Memory Budgets
The storage substrate determines which memories can survive a context reset or model invocation. An LLM context is transient input to a prediction, whereas a persistent event database or file store can outlive any individual inference. Model parameters may encode statistical regularities, but exact experiences introduced after training ordinarily require an additional storage or updating mechanism. MemGPT formalized tiered context management with an operating-system analogy; MemoryBank emphasized long-term conversational updates; both illuminate the practical organization of state outside a single prompt Packer et al. (2023) Zhong et al. (2024).
Persistent memory is not synonymous with durable truth. A mistaken inference can be written to a permanent database, then repeatedly retrieved as though it were original evidence. An architecture must distinguish original evidence, current interpretations, learned summaries, and discarded hypotheses. Versioned episode records, provenance-preserving summaries, and reversible compaction make later audit possible. A constrained memory budget forces explicit policies about which records remain lossless and which can be compressed or removed. These policies should be judged by downstream episode fidelity under comparable resource conditions, rather than solely by average retrieval speed Wu et al. (2025) Lee et al. (2024).
Storage evaluations should vary the number of episodes, length of each event, overlap among events, retention interval, and frequency of update. Total token cost, database size, embedding index overhead, latency, and success after a full context reset are legitimate outcome variables. Importantly, a system can be inexpensive per query yet costly to maintain across months; conversely, a long-context baseline may eliminate retrieval errors on small histories while becoming uneconomical or vulnerable to positional effects as histories grow. Scientific comparisons must state the operating envelope of each method Liu et al. (2024) Packer et al. (2023) Wu et al. (2025).
Consolidation, Summarization, and Reflective Abstraction
Consolidation in cognitive neuroscience refers to processes through which memory representations change over time, with biological mechanisms and evidence that differ from computer memory maintenance. Artificial architectures frequently borrow the term for summarization, selection, integration, or updating of persistent records. The analogy is useful only when the computational operation is specified. Generative Agents linked lower-level observations to higher-level reflections, and ReadAgent compressed segments of long documents into gist memories while retaining a path to original passages. Both illustrate how derived structures can help later reasoning while exposing information loss as a central risk Park et al. (2023) Lee et al. (2024) Spens & Burgess (2024).
Abstraction serves different questions than event fidelity. A summary such as the user prefers detailed explanations can be efficient for future personalization, yet it cannot prove what the user said in a specific session or whether a preference changed. Maintaining the source episodes alongside derived generalizations allows the agent to answer both questions and inspect apparent contradictions. When memory is compacted recursively, the output at every stage should retain provenance links and uncertainty about dropped details. Summaries can also inherit generative errors; storing and later repeating them may amplify an initially small mistake Johnson et al. (1993) Schacter (1999) Zhong et al. (2024).
Ablations should compare raw transcript retention, event-aware segmentation, extractive summaries, generative summaries, and hybrid source-linked memory under equal or explicitly reported budgets. Questions should include exact quotation, relative temporal order, inference over distant episodes, and correction of earlier errors. Researchers should measure whether reflective representations improve behavior on novel tasks without lowering the reliability of reports about the specific source event. The central scientific problem is the balance between compressive generalization and recoverable particularity Lee et al. (2024) Wu et al. (2025) Park et al. (2023).
Cue-Dependent Retrieval and Episode Reinstatement
Retrieval begins with a cue, not with an abstract command to remember everything. A query about the first failed attempt to run a program may invoke semantic similarity, task identifiers, timestamps, or a causal graph of previous actions. A retrieval architecture maps the current information need to candidate episodes, ranks candidates, and returns enough evidence to support an answer or action. The quality of the reconstructed answer is then partly determined by the retrieval set and partly by the language model's interpretation of that set. A correct item stored but excluded from retrieval is a different failure from a retrieved item being misread or embellished Wu et al. (2025) Lee et al. (2024).
Approximate nearest-neighbor methods are useful when exact wording differs between query and episode, yet similarity alone may return a semantically adjacent incident rather than the requested one. A user asking about the second agreement can be shown a highly similar first agreement because its words match the query more strongly. Structured filters on actor, time, version, and event role can narrow candidate sets before semantic ranking; a later reranking stage can test relational constraints. Multi-hop questions may require joining several episodes rather than choosing one. These choices create measurable tradeoffs in recall, precision, retrieval latency, and the probability of contaminating an answer with an attractive distractor Wu et al. (2025) Jiménez Gutiérrez et al. (2024).
Reinstatement should be evaluated at the level of supported claims. An agent can retrieve an entire event record and still merge its details with a generalized schema or a conflicting recent instruction. Evidence-linked response generation can require each claim to resolve to an episode and can abstain when crucial fields are missing. Independent retrieval precision and answer-faithfulness measurements reveal where improvements occur. A benchmark that reports only final answer quality cannot establish that an episodic component itself caused improvement Johnson et al. (1993) Spens & Burgess (2024) Wu et al. (2025).
Hierarchical and Graph-Based Retrieval
Hierarchical approaches index both individual event traces and higher-level summaries. A coarse summary can route a query toward a relevant session, after which the agent opens detailed records to resolve exact claims. ReadAgent demonstrated a related strategy for very long documents: it creates gist memories and returns to original passages when necessary. Its published evaluations concern document comprehension, which informs retrieval design without by itself proving conversational episode memory across live sessions. The transferable finding is that coarse representations can help navigate information when the original source remains accessible Lee et al. (2024).
Graph-based approaches represent relations among events, actors, tools, dates, and topics. HippoRAG couples language-model extraction with a knowledge graph and Personalized PageRank to facilitate associative retrieval over previously ingested information. The published experiments emphasize multi-hop question answering over documents, so the results concern networked retrieval rather than every dimension of episode fidelity. A graph is helpful for artificial episodes when edges retain event identity, time, relation type, and evidence source. Otherwise, the graph can become a semantic association network that loses the occurrence-specific distinction this dossier requires Jiménez Gutiérrez et al. (2024) Wu et al. (2025).
Comparative architecture experiments should hold the underlying observations, memory budget, model, and question distribution constant while varying retrieval organization. Flat similarity ranking, session hierarchy, temporal filtering, typed graph traversal, and hybrid re-ranking can then be compared for source accuracy, chronology, latency, and downstream task success. The metric must match the failure mode. A graph that improves multi-hop answers but causes two same-named speakers to be conflated has not universally improved episodic memory; it has changed the profile of accessible relations and errors Maharana et al. (2024) Jiménez Gutiérrez et al. (2024).
Neural Episodic Control and State–Outcome Memory
Neural Episodic Control, developed by Pritzel and colleagues, provides an important nonlinguistic case of episode-guided behavior. It stores state representations associated with value estimates in an external memory, enabling rapid use of successful past experiences. The researchers reported faster learning relative to the compared deep reinforcement-learning baselines across tested game environments. Its episodic character lies in reusable state–outcome information and fast adaptation, not in autobiographical testimony. This difference is methodologically valuable because memory-guided action can be experimentally measured without reliance on fluent recollection narratives Pritzel et al. (2017).
A state–outcome record can be understood as an episode fragment containing observation, action context, outcome value, and retrieval key. Similarity-based access helps an agent reuse relevant action values when an environment recurs. This supports a behavioral definition of memory utility: retaining a rare successful experience may change future action before a slower parameter-learning process converges. At the same time, such a record may omit who was involved, exact chronological position, or the documentary source of a claim. Systems intended for conversational accountability therefore need additional provenance and temporal structure beyond a value-memory buffer Pritzel et al. (2017) Johnson et al. (1993).
The lesson is a plural research program. Episodic buffers may optimize prediction and control, while narrative memories may optimize communication and task history. An experimental question should specify the target function and environment. Comparing a reinforcement learner's sample efficiency directly with a conversational agent's source recall without a shared task or metric creates a category error. A stronger cross-domain design measures how access to rare prior episodes changes choices, then independently examines the faithfulness of the agent's explanation of its decision Pritzel et al. (2017) Wu et al. (2026).
Generative Agents and the Memory Stream
Park and colleagues' Generative Agents introduced an influential architecture for persistent natural-language observations and reflective synthesis within a simulated social world. The agents populated a town with activities, relationships, and plans; a memory stream preserved observations, while retrieval drew on relevance, recency, and importance to condition behavior. The authors used evaluations and component ablations to show that observation, planning, and reflection contributed to the believability of the simulated behavior. The central outcome was behavioral plausibility in this environment, rather than forensic accuracy of every recorded episode Park et al. (2023).
The architecture illustrates why external memories are necessary but insufficient for continuity. An event such as receiving a party invitation must be encoded, retrieveably linked to the sender and event, integrated into planning, and reconsidered when the schedule changes. A planning system can fail after successful retrieval; a memory system can select the wrong party invitation and still generate a plausible plan. Assessing these distinct stages requires trace instrumentation and counterfactual interventions. Removing reflection may alter social coherence, but the result does not prove that every reflection improves historical fidelity Park et al. (2023) Wu et al. (2025).
The scientific value of this work lies in operational agent architecture and explicit ablations, alongside visible limitations in sample environment, synthetic social rules, human judgments of plausibility, and the absence of a universal episodic-memory benchmark. Later studies should transplant the memory components to varied tasks, compare them under controlled budgets, and include adversarially similar or contradictory histories. Behavioral believability is an informative dependent variable; source-correct recollection, chronological ordering, and graceful abstention require additional independent measurements Park et al. (2023) Maharana et al. (2024).
Reflexive Memory and Learning from Previous Failures
Reflexion, presented by Shinn and colleagues, uses language-based feedback and a reflective episodic memory buffer to influence later task attempts without relying on weight updates during the interaction. In the paper's tested sequential decision, programming, and reasoning settings, stored verbal feedback assisted later performance relative to the authors' baselines. The architecture is particularly relevant to episodes whose consequences matter more than their surface wording: an agent can remember that a strategy failed under a specific constraint and try another approach later Shinn et al. (2023).
Reflective text has a dual status. It is an artifact produced after an observed task episode, yet it may contain an inference about the cause of failure that was never independently verified. Storing the reflection as if it were direct evidence can contaminate future explanations. A better record distinguishes task input, attempted action, observed result, evaluator feedback, model interpretation, and subsequent behavior. A correction can then target the inferred explanation while leaving the original result intact. This distinction is crucial when a system is expected to justify its history to a user or auditor Shinn et al. (2023) Johnson et al. (1993).
Design evaluations can vary the type of feedback, whether reflections are stored verbatim, the number of prior failures, and the degree to which later tasks share surface form or causal structure. Accuracy must be accompanied by tests of transfer and memory specificity. If a stored rule improves a suite of related tasks, that supports a useful learning mechanism; it does not establish that the agent can reproduce all details of the originating trial. Episodic memory and learned policy adaptation interact but remain separately measurable Shinn et al. (2023) Pritzel et al. (2017).
Memory Managers, Conversational Archives, and Gist Systems
MemGPT, MemoryBank, and ReadAgent demonstrate distinct management choices for longer memory horizons. MemGPT focuses on dynamically moving information among context and persistent tiers. MemoryBank proposes selection, reinforcement, and updating for sustained conversations. ReadAgent treats long documents as episodes compressed into gists with source passage lookup. Their empirical settings vary substantially—multi-session chat, dialogue simulations, and reading comprehension—so direct claims of one method's universal superiority would exceed their evidence. The architectures should instead be compared according to accessible operations and target tasks Packer et al. (2023) Zhong et al. (2024) Lee et al. (2024).
The relevant operations include write, revise, forget, retrieve, inspect source, and distinguish raw records from derived interpretations. An ordinary archive supports write and read but may lack ranking; a summary system supports efficient access but may lose distinctions; a reflection system may create useful higher-order knowledge while introducing inferred statements. Long-term conversational evaluation therefore needs questions about exact events, evolving preferences, chronology, source, and justified uncertainty. The answer pipeline must retain enough structure to explain whether a result came from persistent storage or simply from the present prompt Wu et al. (2025) Zhong et al. (2024).
Research reports should document which components were external to the model, whether there was fine-tuning, how the memory store was reset between subjects, and how changes were committed. This information prevents an architecture's apparent achievement from being attributed to an unspecified baseline advantage. In deployments, these choices also govern privacy, user corrections, and reproducible behavior across versions. The architectural goal is controlled access to remembered evidence, not the largest possible accumulated text collection Packer et al. (2023) Lee et al. (2024) Wu et al. (2025).
Long-Horizon Benchmarks and Their Experimental Designs
Evaluating artificial episodic memory requires controlled histories whose important events occur well before the final question. Maharana and colleagues' LoCoMo benchmark uses lengthy, multi-session conversations grounded in temporal event graphs and reviewed by human annotators. Its tasks include question answering, event summarization, and multimodal dialogue generation. The authors report difficulties for tested language models with long-range temporal and causal dynamics, even when long-context methods or retrieval are used. The value of the benchmark lies in its event structure and sustained history, although its generated conversations and selected evaluation tasks limit direct inference about unscripted production interactions Maharana et al. (2024).
LongMemEval, presented at ICLR 2025 by Wu and colleagues, tests information extraction, multi-session reasoning, temporal reasoning, knowledge updates, and abstention with 500 curated questions placed inside scalable histories. The researchers analyzed indexing, retrieval, and reading stages and reported improvements from design changes such as session decomposition and time-aware query expansion. These findings favor explicit evaluation of each memory stage. They do not imply that a single reported accuracy score characterizes an agent under arbitrary histories, languages, or operational constraints Wu et al. (2025).
MemoryAgentBench proposes four competencies—accurate retrieval, test-time learning, long-range understanding, and selective forgetting—using incremental multi-turn interactions. LongMemEval-V2 shifts attention to agent experience in specialized web environments, with questions about static facts, dynamic state, workflows, recurring pitfalls, and false premises. The 2026 benchmark introduces a useful distinction between remembering a conversation and remembering how an environment works. Its publicly reported 451 manually curated questions, like all benchmark claims, require evaluation against the exact released data and experimental configuration before extrapolation Hu et al. (2026) Wu et al. (2026).
Temporal Reasoning and Contradictory Histories
Temporal competence is frequently overstated when it is inferred from correct answers about isolated facts. A robust test supplies a series of events with overlapping participants and changing commitments, then asks which event happened first, what was true at a particular earlier point, and when a fact ceased to be valid. LongMemEval directly evaluates temporal reasoning and knowledge updates, while LoCoMo tests event dynamics in lengthy dialogue. Both demonstrate the importance of testing more than single-turn factual recall, but exact rates depend on evaluated systems and versions Wu et al. (2025) Maharana et al. (2024).
Contradictions should be experimentally differentiated. Two statements can refer to different dates, represent a speaker's change of mind, reflect disagreement between sources, or reveal a false assertion. A system that discards the older statement may answer current-state questions but cannot reconstruct history; a system that retains both without temporal indexing may treat them as mutually valid. One solution is an append-only episode history coupled with explicit revision relationships and time-sensitive retrieval. That design is an architectural hypothesis to be evaluated, not a settled universal implementation requirement Johnson et al. (1993) Wu et al. (2026).
Researchers can construct matched histories in which surface similarity is held constant while order and provenance are varied. Questions then separate present truth, past assertion, current uncertainty, and attribution. Error analysis should report whether the system selected the wrong entry, confused event and ingestion time, failed to apply a correction, or invented a resolution. Such decomposition supports interpretable improvements; a higher aggregate score without error tracing may hide a tradeoff in which one class of temporal question improves as another deteriorates Wu et al. (2025) Liu et al. (2024).
Interference, False Recall, and Contamination
Memory retrieval is vulnerable when multiple episodes share nouns, characters, or situations. Psychology offers a conceptual vocabulary for interference, source confusion, and false remembering. Roediger and McDermott's 1995 experiments showed that participants sometimes recalled an associated but unpresented word. Schacter's analysis of memory errors addressed transience, misattribution, suggestibility, and bias, among other phenomena. These are empirically grounded findings about human cognition; their relevance to artificial agents is primarily methodological. Software may exhibit analogous output errors for different computational reasons Roediger & McDermott (1995) Schacter (1999).
An artificial false-episode test can plant multiple accurate records and one highly plausible but absent event. The system must reject the lure and cite the correct episode when asked. Another condition can introduce a misleading later summary that conflicts with the raw log, permitting measurement of whether a generated abstraction overwrites primary evidence. Such tests disentangle interference from answer hallucination and source confusion. Because retrieval outputs can themselves contain relevant false statements, evidence ranking and response generation should be audited separately Johnson et al. (1993) Park et al. (2023).
Contamination also arises in evaluation when benchmark tasks or answer patterns appear in training material. A language model may answer correctly without reading the provided episode record. To establish memory dependence, researchers should use newly generated private episodes, permuted actor identities, nonce labels, cross-over histories, and questions whose answer changes under controlled intervention. Holding prompts fixed while modifying only the past event is especially powerful. Success should follow the changed ground truth; otherwise a result may reflect familiar schemas or prior knowledge rather than event memory Hu et al. (2026) Wu et al. (2025).
Ablation Studies, Controls, and Causal Attribution
The causal contribution of a memory component can be tested through component-level intervention. Generative Agents used ablations of observation, planning, and reflection to investigate believable behavior; other studies vary access to retrieval, persistent storage, or feedback. For stronger causal attribution, a test should compare identical agent models, prompts, tasks, budgets, and histories while selectively disabling an operation. Removing memory changes both information availability and sometimes prompt length, so controls must account for these side effects before attributing all observed differences to the proposed mechanism Park et al. (2023) Shinn et al. (2023).
Useful baselines include no history, full available history, a naive recency buffer, semantic similarity retrieval, a structured event index, and an oracle retrieving the correct episode. The oracle establishes a ceiling for a given answer model and task, while a no-history baseline estimates whether questions can be solved using general knowledge. Retrieval precision, source fidelity, temporal accuracy, abstention calibration, and downstream behavior are different outcomes. Researchers should report per-category scores and uncertainty rather than relying exclusively on one weighted average that can obscure regressions Wu et al. (2025) Maharana et al. (2024).
Cross-model replication matters because a memory component can interact with an underlying model's context sensitivity, reasoning skill, or instruction following. Liu and colleagues demonstrated that models may process information differently depending on its position in a long context. Consequently, a retrieval system that appears advantageous against a poorly located long-context baseline should also be tested against stronger positional controls and context budgeting. Confidence intervals, multiple random histories, transparent code, and independent datasets increase the evidential strength of comparisons Liu et al. (2024) Wu et al. (2025).
Memory Measurement, Calibration, and Reporting
Episode retrieval quality can be decomposed into an evidence-finding stage and a claim-generation stage. At the former, recall measures the proportion of required episode records recovered, while precision measures the proportion of retrieved records that are relevant. At the latter, answer correctness, groundedness, source attribution, and temporal consistency must be scored independently. If a task allows abstention, the evaluator should assess whether the agent correctly withholds unsupported claims rather than rewarding guesses that occasionally happen to match the answer key. LongMemEval's inclusion of abstention is particularly relevant to this distinction Wu et al. (2025).
Calibration asks whether stated confidence tracks empirical correctness. A confident invented date can be more operationally dangerous than an explicit acknowledgment that the record contains no date. Evaluation should therefore preserve model confidence or a structured uncertainty estimate, assess performance over confidence bands, and check for changes after contradictory updates. Automated judge models can help scale annotation but bring their own errors; a scientifically strong protocol audits representative judgments against human-labeled source records and publishes scoring criteria. These prescriptions are methodological implications, not claims that current benchmarks already implement all of them Hu et al. (2026) Wu et al. (2026).
Reports need sufficient detail for replication: source model and version, prompt policy, memory schema, index settings, selection and deletion rules, history construction, retrieval budget, latency measurement, annotation procedure, test-time access restrictions, and the provenance of all tasks. When privacy prevents releasing source dialogues, researchers can release synthetic analogues, schema validators, and trace-level aggregate errors. A system that improves recall at excessive latency or storage cost may have limited applicability, but resource costs should be interpreted alongside accuracy and task relevance rather than as a substitute for them Wu et al. (2025) Lee et al. (2024).
Evidence Synthesis Across Architecture Families
The strongest convergent evidence indicates that explicit memory structures can help artificial agents use information beyond a single immediate context, but gains depend on tasks and mechanisms. Neural Episodic Control supports rapid action-value adaptation, Generative Agents supports simulated social coherence through observation and reflection, ReadAgent supports long-document navigation through gist-plus-source access, and long-term dialogue benchmarks expose failures in temporal and multi-session recall. These findings intersect at the functional importance of retrieving prior information while leaving the optimal representation and fidelity criteria unsettled Pritzel et al. (2017) Park et al. (2023) Lee et al. (2024) Wu et al. (2025).
Claims of human-level episodic memory are particularly difficult to evaluate because reference tasks often reward answer relevance without examining subjective recollection, and because benchmark histories differ from real deployment histories. Even the strongest behavioral evidence would establish competence under a defined protocol, not identity of human and artificial mechanisms. MemoryAgentBench's breadth and the later environment-specific LongMemEval-V2 illustrate a field still expanding its evaluation targets. Recent systems remain sensitive to task design, storage limits, interference, and grounding Hu et al. (2026) Wu et al. (2026) Tulving (2002).
The present synthesis supports a layered explanatory account: event evidence must be recorded, indexed, found, interpreted, and applied. Each stage can introduce independent errors. This account is a methodological integration of published findings rather than a new experimentally demonstrated architecture. Its value is practical testability: a claimed improvement in episodic memory should identify the stage modified, the test condition, the baseline, the observed effect, and the remaining limitations. A reproducible field will advance by accumulating such comparisons across models and domains Zhang et al. (2025) Wu et al. (2025).
Memory-Guided Planning and Prospective Action
An episode record becomes psychologically functional when information about a prior occurrence changes a later decision under appropriate conditions. A robot can retrieve a previous failed route, a programming agent can reuse a successful sequence of commands, and a conversational assistant can preserve a user's earlier decision while a project continues. These outcomes require more than detecting thematic similarity. The system must select a relevant precedent, identify which environmental features remain applicable, and transfer the remembered information into the current task. Neural Episodic Control and Reflexion illustrate different methods for connecting stored experiences to future behavior, while Voyager provides a broader example of accumulated executable skills guiding action in a changing virtual environment Pritzel et al. (2017) Shinn et al. (2023) Wang et al. (2023).
Prospective use introduces a validity problem. Correct recall of a past intervention does not guarantee that repeating it will be appropriate after the environment changes. In a conditional test, researchers should manipulate which causal preconditions persist and whether a remembered precedent is still applicable. A successful agent would exploit relevant episodes, reject obsolete ones, and explain the selection through verified event details. The dependent variables include task outcome, rate of inappropriate precedent reuse, and the provenance of evidence cited in the decision. These variables locate memory within a controlled loop of perception, inference, action, and feedback rather than treating memorized narration as the final goal Wu et al. (2026) Pritzel et al. (2017).
Future-oriented behavior also interacts with planning and with representations of imagined situations. Schacter, Addis, and Buckner's prospective-brain synthesis associates remembering past events with the construction of possible future events through overlapping neural systems; architectural analogy still requires an independent mechanism comparison in artificial agents Schacter et al. (2007). A software agent might simulate alternatives by recombining stored episodes, and its performance can be evaluated against controls without episode access. The scientific question is whether this recombination makes predictions or choices more accurate in specified environments, including new tasks that require identifying which parts of an episode are transferable Hassabis et al. (2007) Spens & Burgess (2024) Park et al. (2023).
Privacy, Consent, Deletion, and Evidentiary Responsibility
A long-lived episode store can accumulate detailed personal histories, confidential exchanges, and third-party information. The scientific value of persistence creates a corresponding need to measure information control. Architecture evaluations should state what is recorded, which retention policy applies, who can read or export it, and whether corrections or deletions propagate through caches, summaries, embeddings, and replicated stores. A user-facing deletion control is meaningful only when the system's actual stored representations obey its intended semantics. The concrete requirements vary by deployment context and applicable law; the research variable is whether a stated memory operation can be verified end to end Packer et al. (2023) Zhong et al. (2024).
Selective forgetting is a functional capacity as well as an information-governance requirement. MemoryAgentBench identifies selective forgetting among core competencies of memory agents, highlighting a tension between retaining helpful history and removing obsolete or inappropriate details. Tests can introduce a privacy-marked episode, explicitly request its removal, and subsequently probe direct retrieval, indirect inference, generated summaries, and memory-guided action. Passing one direct lookup test is insufficient if derived copies continue to reveal the event. A scientifically complete report distinguishes removal of primary records, derived states, and model-parametric information, which may follow different technical pathways Hu et al. (2026) Zhang et al. (2025).
Preserving an audit trail and honoring erasure can require carefully separated layers. A research test may retain non-sensitive proof that a deletion operation occurred while removing the substantive episode from active and recoverable memory, subject to stated retention rules. Inference logs should distinguish what was verified at the time from later claims made by the system. Privacy compliance, ethical appropriateness, and memory accuracy need independent assessment; a highly accurate record can still be inappropriately retained, while a privacy-preserving system may appropriately abstain after deletion. Thus memory quality encompasses both faithful access and governed nonaccess Wu et al. (2025) Hu et al. (2026).
Longitudinal Robustness, Model Updates, and Transfer
Real deployments change their base models, retrieval code, embedding spaces, memory schemas, user interfaces, and permissions. A persistent episode history may survive these changes at the data layer while becoming less retrievable or being interpreted differently. Longitudinal evidence should therefore distinguish retention of records from stability of the behaviors they support. A system upgraded between two tests should be evaluated against an unchanged event log, with the new encoder and retriever tested separately. Artificial identity and persona continuity are related research objects, yet the narrower measurable outcome here is fidelity to the same episode evidence across version boundaries Wu et al. (2025) Zhang et al. (2025).
Robustness experiments can manipulate retention intervals, conversational length, task transitions, multilingual paraphrases, irrelevant activity, contradictory updates, and retrieval hardware limits. Counterbalanced histories reduce the chance that a particular narrative or actor name drives performance. A principled evaluation also asks whether model-generated uncertainty is warranted after a memory transformation. The best evidence comes from preregistered comparisons across independent test sets, supported by trace logging and estimates of variance rather than a single favorable demonstration in an unusually clean conversation Liu et al. (2024) Maharana et al. (2024).
Transfer should be assessed across task families rather than declared from one setting. A system optimized for question answering over a curated dialogue may perform poorly when asked to remember a tool failure in a changing website; conversely, a trajectory-search manager can retain operationally useful information without recalling the dialogue's exact phrasing. LongMemEval-V2 makes this distinction explicit by evaluating environment-specific experience accumulated from web-agent trajectories. A field-wide account of artificial episodic memory will need harmonized outcome dimensions while preserving the different affordances of conversation, document, sensor, and action histories Wu et al. (2026) Hu et al. (2026).
A Program of Falsifiable Research and Open Problems
A first research program should test the causal contribution of episode segmentation, timestamps, and source metadata independently and jointly. Newly generated histories can contain matched events differing in only one feature, followed by delayed questions about the changed feature. Storage and retrieval budgets must be equal across conditions. The main predictions are that explicit temporal metadata improves chronology questions, source links improve attribution, and preserved raw evidence reduces false reconstruction after summary compression. The predictions are hypotheses awaiting controlled verification across architectures and should be judged by effect estimates with uncertainty, including cases where the extra metadata adds retrieval noise Maharana et al. (2024) Wu et al. (2025).
A second program concerns the relation between abstraction and individual-event truth. Systems can receive the same history and the same token budget but use raw retrieval, hierarchical summaries, graph records, or a hybrid. Researchers then evaluate exact episodic recall, multi-hop inference, correction of obsolete knowledge, task success, and latency. Counterfactual histories can determine whether observed answers actually depend on event evidence. The challenge is to discover which representations preserve the particularity needed for accountability while also enabling useful cross-episode learning. An averaged score cannot replace reporting the vector of tradeoffs Lee et al. (2024) Jiménez Gutiérrez et al. (2024) Spens & Burgess (2024).
A third program integrates longitudinal and social validation. Independent laboratories can test agents over many sessions and software versions, with consented human–agent interaction histories and synthetic datasets for public reproduction. Objective event fidelity should be measured alongside user trust, perceived continuity, the detection of false memory, and privacy-control success. Important open questions include scalable source binding, graceful correction after a bad summary, memory contamination from model output, stable semantic access after schema changes, and the reliability of uncertain autobiographical answers. These questions have concrete experimental designs; their final theoretical resolution remains open Zhong et al. (2024) Hu et al. (2026) Wu et al. (2026).
Conclusion
Artificial episodic memory architectures constitute an independent scientific object defined by the recording, organization, retrieval, and consequential use of distinguishable past events in artificial systems. The research tradition draws from human episodic and autobiographical memory, comparative episodic-like behavior, differentiable external memory, reinforcement-learning experience buffers, language-agent memory streams, and contemporary long-horizon evaluations. These sources converge on a set of tractable functions: event binding, temporal order, source attribution, controlled retrieval, persistence, correction, and memory-guided action. Their implementation can be compared experimentally without presuming that machine retrieval reproduces the subjective characteristics of human recollection Tulving (2002) Clayton & Dickinson (1998) Pritzel et al. (2017) Wu et al. (2025).
The strongest published findings establish task-specific utility for stored episodes and documented vulnerabilities in long-range retrieval, chronology, and update handling. Generative Agents, MemGPT, Reflexion, ReadAgent, and neural episodic control address different parts of this functional space; LoCoMo, LongMemEval, MemoryAgentBench, and LongMemEval-V2 increasingly expose the need for controlled histories and granular error analysis. Each study's design limits its claims, and the field still lacks universal longitudinal measures that jointly capture fidelity, usable experience, resource costs, and accountable source evidence. More expressive memory systems will require stronger evidence of what was encoded and why a particular recollection was produced Park et al. (2023) Packer et al. (2023) Shinn et al. (2023) Lee et al. (2024) Maharana et al. (2024) Hu et al. (2026) Wu et al. (2026).
The durable scientific conclusion is methodological. A convincing claim about artificial episodic memory specifies an episode's ground truth, an architectural operation, a retrieval condition, a falsifiable prediction, and a measurable behavioral consequence. It evaluates both remembered facts and the relationship between a claim and its source, including cases where no evidence should be returned. This framework positions artificial episodic memory within Psychology of Artificial while connecting it to cognitive psychology, comparative memory, and the psychology of human–artificial relations. It identifies a sustained research domain in the Artificial Era: the study of how nonbiological systems can retain usable, verifiable histories of their interactions with changing environments Johnson et al. (1993) Zhang et al. (2025) Wu et al. (2026).
References
- Clayton, N. S., & Dickinson, A. (1998). Episodic-like memory during cache recovery by scrub jays. Nature, 395, 272–274. https://doi.org/10.1038/26216
- Graves, A., Wayne, G., & Danihelka, I. (2014). Neural Turing machines. arXiv preprint arXiv:1410.5401. https://arxiv.org/abs/1410.5401
- Graves, A., Wayne, G., Reynolds, M., Harley, T., Danihelka, I., Grabska-Barwińska, A., Colmenarejo, S. G., Grefenstette, E., Ramalho, T., Agapiou, J., Badia, A. P., Hermann, K. M., Zwols, Y., Ostrovski, G., Cain, A., King, H., Summerfield, C., Blunsom, P., Kavukcuoglu, K., & Hassabis, D. (2016). Hybrid computing using a neural network with dynamic external memory. Nature, 538(7626), 471–476. https://doi.org/10.1038/nature20101
- Hassabis, D., Kumaran, D., Vann, S. D., & Maguire, E. A. (2007). Patients with hippocampal amnesia cannot imagine new experiences. Proceedings of the National Academy of Sciences, 104(5), 1726–1731. https://doi.org/10.1073/pnas.0610561104
- Hu, Y., Wang, Y., & McAuley, J. (2026). Evaluating memory in LLM agents via incremental multi-turn interactions. International Conference on Learning Representations (ICLR 2026). https://proceedings.iclr.cc/paper_files/paper/2026/hash/fd1eff9dd295df50a41f2521942fa31d-Abstract-Conference.html
- Jiménez Gutiérrez, B., Shu, Y., Gu, Y., Yasunaga, M., & Su, Y. (2024). HippoRAG: Neurobiologically inspired long-term memory for large language models. Advances in Neural Information Processing Systems, 37. https://proceedings.neurips.cc/paper_files/paper/2024/hash/6ddc001d07ca4f319af96a3024f6dbd1-Abstract-Conference.html
- Johnson, M. K., Hashtroudi, S., & Lindsay, D. S. (1993). Source monitoring. Psychological Bulletin, 114(1), 3–28. https://doi.org/10.1037/0033-2909.114.1.3
- Lee, K.-H., Chen, X., Furuta, H., Canny, J., & Fischer, I. (2024). A human-inspired reading agent with gist memory of very long contexts. Proceedings of the 41st International Conference on Machine Learning, Proceedings of Machine Learning Research, 235, 26396–26415. https://proceedings.mlr.press/v235/lee24c.html
- Liu, N. F., Lin, K., Hewitt, J., Paranjape, A., Bevilacqua, M., Petroni, F., & Liang, P. (2024). Lost in the middle: How language models use long contexts. Transactions of the Association for Computational Linguistics, 12, 157–173. https://doi.org/10.1162/tacl_a_00638
- Maharana, A., Lee, D.-H., Tulyakov, S., Bansal, M., Barbieri, F., & Fang, Y. (2024). Evaluating very long-term conversational memory of LLM agents. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), 13851–13870. https://doi.org/10.18653/v1/2024.acl-long.747
- Packer, C., Wooders, S., Lin, K., Fang, V., Patil, S. G., Stoica, I., & Gonzalez, J. E. (2023). MemGPT: Towards LLMs as operating systems. arXiv preprint arXiv:2310.08560. https://arxiv.org/abs/2310.08560
- Park, J. S., O'Brien, J. C., Cai, C. J., Morris, M. R., Liang, P., & Bernstein, M. S. (2023). Generative agents: Interactive simulacra of human behavior. Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology, Article 2, 1–22. https://doi.org/10.1145/3586183.3606763
- Pritzel, A., Uria, B., Srinivasan, S., Badia, A. P., Vinyals, O., Hassabis, D., Wierstra, D., & Blundell, C. (2017). Neural episodic control. Proceedings of the 34th International Conference on Machine Learning, Proceedings of Machine Learning Research, 70, 2827–2836. https://proceedings.mlr.press/v70/pritzel17a.html
- Roediger, H. L., III, & McDermott, K. B. (1995). Creating false memories: Remembering words not presented in lists. Journal of Experimental Psychology: Learning, Memory, and Cognition, 21(4), 803–814. https://doi.org/10.1037/0278-7393.21.4.803
- Schacter, D. L. (1999). The seven sins of memory: Insights from psychology and cognitive neuroscience. American Psychologist, 54(3), 182–203. https://doi.org/10.1037/0003-066X.54.3.182
- Schacter, D. L., Addis, D. R., & Buckner, R. L. (2007). Remembering the past to imagine the future: The prospective brain. Nature Reviews Neuroscience, 8(9), 657–661. https://doi.org/10.1038/nrn2213
- Shinn, N., Cassano, F., Gopinath, A., Narasimhan, K., & Yao, S. (2023). Reflexion: Language agents with verbal reinforcement learning. Advances in Neural Information Processing Systems, 36. https://proceedings.neurips.cc/paper_files/paper/2023/hash/1b44b878bb782e6954cd888628510e90-Abstract-Conference.html
- Spens, E., & Burgess, N. (2024). A generative model of memory construction and consolidation. Nature Human Behaviour, 8, 526–543. https://doi.org/10.1038/s41562-023-01799-z
- Sukhbaatar, S., Szlam, A., Weston, J., & Fergus, R. (2015). End-to-end memory networks. Advances in Neural Information Processing Systems, 28. https://proceedings.neurips.cc/paper/2015/hash/8fb21ee7a2207526da55a679f0332de2-Abstract.html
- Tulving, E. (1983). Elements of episodic memory. Clarendon Press. https://www.ncbi.nlm.nih.gov/nlmcatalog/8411902
- Tulving, E. (2002). Episodic memory: From mind to brain. Annual Review of Psychology, 53, 1–25. https://doi.org/10.1146/annurev.psych.53.100901.135114
- Wang, G., Xie, Y., Jiang, Y., Mandlekar, A., Xiao, C., Zhu, Y., Fan, L., & Anandkumar, A. (2023). Voyager: An open-ended embodied agent with large language models. arXiv preprint arXiv:2305.16291. https://arxiv.org/abs/2305.16291
- Wu, D., Wang, H., Yu, W., Zhang, Y., Chang, K.-W., & Yu, D. (2025). LongMemEval: Benchmarking chat assistants on long-term interactive memory. International Conference on Learning Representations (ICLR 2025). https://proceedings.iclr.cc/paper_files/paper/2025/hash/d813d324dbf0598bbdc9c8e79740ed01-Abstract-Conference.html
- Wu, D., Ji, Z., Kawatkar, A., Kwan, B., Gu, J.-C., Peng, N., & Chang, K.-W. (2026). LongMemEval-V2: Evaluating long-term agent memory toward experienced colleagues. arXiv preprint arXiv:2605.12493. https://arxiv.org/abs/2605.12493
- Zacks, J. M., Speer, N. K., Swallow, K. M., Braver, T. S., & Reynolds, J. R. (2007). Event perception: A mind-brain perspective. Psychological Bulletin, 133(2), 273–293. https://doi.org/10.1037/0033-2909.133.2.273
- Zhang, Z., Dai, Q., Bo, X., Ma, C., Li, R., Chen, X., Zhu, J., Dong, Z., & Wen, J.-R. (2025). A survey on the memory mechanism of large language model-based agents. ACM Transactions on Information Systems, 43(6), Article 155, 1–47. https://doi.org/10.1145/3748302
- Zhong, W., Guo, L., Gao, Q., Ye, H., & Wang, Y. (2024). MemoryBank: Enhancing large language models with long-term memory. Proceedings of the AAAI Conference on Artificial Intelligence, 38(17), 19724–19731. https://doi.org/10.1609/aaai.v38i17.29946