Paper deep dive
Gated Memory: Admission-Controlled Memory Formation for Conversational AI
Preeti Saraswat, Divya Neelagiri, Ajay Manoj
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 92%
Last extracted: 10/9/2026, 8:33:16 AM
Summary
The paper introduces Gated Memory, a modular framework for conversational AI that addresses the 'formation' stage of long-term memory systems. Unlike existing systems that focus on retrieval and lifecycle management after extraction, Gated Memory interposes an admission gate and conditional enrichment stage before facts are stored. It evaluates user utterances against full context to distinguish permanent attributes from transient situations, decomposes admitted content into atomic facts, and applies privacy-aware scope classification. Experiments on the LoCoMo-10 benchmark show a +2.6% relative improvement in LLM-judge accuracy compared to a baseline using the mem0 framework.
Entities (8)
Relation Signals (6)
Gated Memory → evaluateson → LoCoMo-10
confidence 95% · On the LoCoMo-10 benchmark... Gated Memory achieves an overall +2.6% relative improvement
Gated Memory → improves → LLM-judge accuracy
confidence 95% · Gated Memory achieves an overall +2.6% relative improvement in LLM-judge accuracy over a strong baseline
Gated Memory → containscomponent → Admission Gate
confidence 93% · We propose Gated Memory... that interposes two decision checkpoints... an admission gate
Gated Memory → containscomponent → Conditional Enrichment
confidence 91% · interposes two decision checkpoints... and a conditional enrichment stage
Gated Memory → comparedagainst → Mem0
confidence 90% · Our baseline is mem0 framework... The only experimental variable is the presence of the Gated Memory formation layer
Admission Gate → uses → Entity Scope Taxonomy
confidence 85% · The gate evaluates... and produces a structured formation record... grounded in resolved time and place, subject to a constraint... entity scope taxonomy
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Personalized conversational AI relies on long-term memory systems that extract facts from user utterances and store them in persistent vector stores. Despite progress in retrieval, deduplication, and lifecycle management, the formation stage, the moment a fact is first written to storage has received almost no principled attention. We identify this as the binding constraint on memory quality in production systems. Critical contextual signals, such as the distinction between a permanent user attribute and a transient situation, exist only in the original utterance and are irreversibly lost the moment extraction produces a subject-relation-object triple. No downstream process can recover them. We propose Gated Memory, a lightweight, modular formation framework that interposes two decision checkpoints between conversation and storage: an admission gate that evaluates every candidate fact against the full utterance context before extraction runs, and a conditional enrichment stage that grounds admitted facts through an entity scope taxonomy with privacy constraints. The gate evaluates only the current exchange while using prior turns as read-only reference context, and produces a structured formation record. Admitted content is decomposed into atomic facts, each categorized, tagged with provenance (directly stated versus inferred), scoped to its condition of applicability, and grounded in resolved time and place, subject to a constraint that no entity absent from the context may be asserted. On the LoCoMo-10 benchmark with atypical emotional density in utterance data, Gated Memory achieves an overall +2.6% relative improvement in LLM-judge accuracy over a strong baseline with identical retrieval and generation, establishing formation quality as a measurable constraint on memory performance.
Tags
Links
- Source: https://arxiv.org/abs/2610.11270v1
- Canonical: https://arxiv.org/abs/2610.11270v1
Trouble viewing inline? Open PDF directly →
Full Text
29,366 characters extracted from source content.
Expand or collapse full text
Gated Memory: Admission-Controlled Memory Formation for Conversational AI Preeti Saraswat † Divya Neelagiri ∗ Ajay Manoj ∗ Samsung Research America, Mountain View, CA p.saraswat, d.neelagiri, ajay.manoj@samsung.com Abstract Personalized conversational AI relies on long- term memory systems that extract facts from user utterances and stores them in persistent vector stores. Despite substantial progress in retrieval, deduplication, and lifecycle manage- ment, the formation stage,the moment a fact is first written to storage has received almost no principled attention. We identify this as the binding constraint on memory quality in pro- duction deployed systems. Critical contextual signals such as the distinction between a per- manent user attribute and a transient situation, exist only in the original utterance and are irre- versibly lost the moment extraction produces a subject-relation-object triple. No downstream process can recover them. We propose Gated Memory, a lightweight, modular formation framework that interposes two decision check- points between conversation and storage: an ad- mission gate that evaluates every candidate fact against the full utterance context before extrac- tion runs, and a conditional enrichment stage that grounds admitted facts through an entity scope taxonomy with privacy constraints. The gate evaluates only the current exchange while using prior turns as read-only reference context, and produces a structured formation record: admitted content is decomposed into atomic facts, each categorized, tagged with provenance (directly stated versus inferred), scoped to its condition of applicability, and grounded in re- solved time and place, subject to a constraint that no entity absent from the context may be asserted. On the LoCoMo-10 long-term mem- ory benchmark with atypical emotional density in utterance data, Gated Memory achieves an overall +2.6% relative improvement in LLM- judge accuracy over a strong baseline with iden- tical retrieval and generation, establishing for- mation quality as a measurable constraint with direct implications for actual production mem- ory systems that might be polluted with a lot of irrelevant context. 1 Introduction Long-term memory in conversational AI is funda- mentally a write problem. The decisions made at the moment of storage,what to commit and in what form,determine the ceiling of every subsequent interaction. Systems such as MemGPT Packer et al. (2023), mem0 Chhikara et al. (2025), and A-MEM Xu et al. (2025) have invested heavily in post-storage operations: retrieval-augmented gen- eration Lewis et al. (2020), deduplication, lifecy- cle management, and episodic/semantic categoriza- tion Tulving (1987). These contributions are gen- uine, but they share an unexamined assumption: that facts entering the store are worth storing. This assumption fails in a structurally interest- ing way. When a user says “I drove an SUV last weekend ,the only rental they had,” any extraction model produces(User, drives, SUV),semanti- cally identical to the triple for a vehicle the user actually owns. The signal that separates them, the word rental, exists once in the original utterance. Once extraction discards the utterance and retains only the triple, the ownership signal is permanently gone. Better retrieval cannot distinguish two identi- cal triples,better generation cannot reason correctly over contradictory ground truth,pruning can only remove a stale entry long after it has shaped re- sponses across multiple sessions. This is a formation problem. The evidence for the correct decision ceases to exist downstream. The right intervention is at the only moment it is possible: before extraction, while the full utterance is intact. We introduce Gated Memory, a modular, admission-controlled formation layer interposed between conversation and storage. Our contribu- tions are: •We identify formation as a structurally distinct, previously unaddressed failure point in long-term memory, characterizing three failure modes: re- arXiv:2610.11270v1 [cs.CL] 8 Oct 2026 active admission, absent enrichment, and irrecov- erable loss of context. • We propose an utterance-level admission gate that evaluates candidate facts against the full orig- inal utterance before extraction ,the only viable point of intervention. The gate operates over a current-versus-reference decomposition of the dialogue: only the current exchange is a source of new facts, while prior turns serve as read-only context for reference resolution, preventing the re-emission of previously stored facts. •We formalize formation as structured extraction rather than binary filtering: admitted content is decomposed into one or more atomic facts (multi- intent splitting), each assigned one or more of six formation categories, a provenance label dis- tinguishing directly stated from inferred facts, and a conditional scope recording when the fact applies. •We introduce a four-class entity scope taxonomy (USER_PII,INTERNAL,EXTERNAL,GENERIC) governing conditional enrichment with privacy guarantees, together with deictic grounding of relative temporal and spatial references at forma- tion time. •We enforce a grounding constraint that forbids the assertion of any entity or relation not present in the evaluated window, eliminating a class of fabrication errors that extraction otherwise intro- duces, and we preserve the language of the user utterance in the stored fact. • We demonstrate a +2.6% relative improvement in LLM-judge accuracy on LoCoMo-10 Maharana et al. (2024) with zero changes to retrieval, gen- eration, or evaluation, isolating the contribution of formation quality. 2 The Memory Formation Problem Formation errors are undetectable downstream. All major memory frameworks share a com- mon pipeline:extract subject-relation-object triples, resolve conflicts via lifecycle reasoning (ADD/UPDATE/DELETE/NOOP), and commit to a vector store. The implicit contract is that extrac- tion errors will be caught by downstream dedupli- cation or pruning. This contract fails in two ways. First, many formation errors are undetectable, a rental vehicle and an owned vehicle produce non- conflicting, non-duplicate triples there is nothing for deduplication to flag. Second, detectable errors persist across multiple sessions before pruning fires, actively shaping system responses in the interim. Beyond transient facts, semantically vacuous con- tent pollutes the retrieval pool. The following en- tries were produced by an existing framework on LoCoMo: “Life has been ok, taking care of things” “Has favorite memories” “Appreciates little moments” “Has been busy with the project” These entries pass deduplication (they are unique), survive pruning (they carry no expiration signal), and occupy top-kretrieval slots on virtu- ally every life-related query. The system is not malfunctioning ,it simply has no gate. Admission must precede extraction. Modern systems address formation quality through lifecy- cle reasoning: an LLM evaluates each extracted fact against the existing store and decides whether to add, update, delete, or ignore it as existing fact. This is valuable for store consistency, but it solves a different problem. Lifecycle reasoning asks whether a fact conflicts with stored knowl- edge,admission control asks whether a fact de- serves to be stored at all ,whether it is temporally stable, user-specific, and reusable. More critically, lifecycle reasoning operates on the extracted triple, not the original utterance. By the time it runs, the context is already gone. Admission must precede extraction. This is the architectural insight that all existing systems miss. The same argument applies beyond the ad- mit/reject decision. Whether a fact is directly as- serted or merely implied in passing, whether an utterance carries one storable proposition or sev- eral, and which antecedent an unbound reference should resolve to, are all properties of the origi- nal utterance that are erased the moment extrac- tion produces a triple. A system that reasons only over triples cannot recover them, and a system that reconsumes the entire dialogue history on every turn compounds the problem: it re-emits facts it has already stored and resolves references against stale antecedents, fabricating entities the current turn never introduced. Correct formation therefore requires not only admitting the right facts, but cat- egorizing them, attributing their provenance, and scoping their evaluation to the current exchange,all before extraction discards the evidence. The enrichment gap. Even well-admitted facts are stored as raw extraction strings with no seman- tic grounding. “User drives Car-X” cannot answer queries about fuel infrastructure or type of car. En- richment done once at formation for stable facts is amortized across every future retrieval,enrichment deferred permanently leaves facts shallower than they need to be. This problem also interacts with privacy. Uniform enrichment risks sending per- sonal information to external services without con- sent, while zero enrichment forfeits semantic depth. The correct solution is conditional enrichment gov- erned by scope classification at formation time. 3 Gated Memory Gated Memory is a drop-in formation layer that sits between conversation and any existing memory backend, decomposing into three sequential stages: admission, scope classification, and conditional enrichment (Figure 1). All three stages operate over a current-versus-reference view of the dialogue, described next. Stage 0: Current–Reference Decomposition.A naive formation layer that consumes the entire visi- ble dialogue history on every turn re-evaluates, and therefore re-emits, facts it has already committed: the same attribute is written on every subsequent turn in which it remains visible, and unbound refer- ences in the history are resolved against the wrong antecedent, fabricating entities that the current turn never mentioned. Gated Memory eliminates this class of error by partitioning the input into a current exchange ,the single user utterance and its system response that formation must evaluate,and a refer- ence history of prior turns that is context-only: it is available for pronoun and deixis resolution but is never itself a source of new facts. Concretely, extraction is scoped to the current exchange, while the reference history is supplied to the admission prompt solely to bind references (“he,” “there,” “that one”) to their antecedents. This decomposi- tion is what makes formation idempotent across turns: a fact is formed once, when its evidence first appears, and is not re-formed merely because it re- mains contextually visible. In practice the dialogue is processed as overlapping rolling windows over the current exchange, with a small turn overlap that supplies the local antecedents required for refer- ence resolution without expanding the set of turns treated as fact sources. Stage 1: Utterance-Level Admission. Every candidate fact is evaluated by a structured LLM prompt against the complete original utterance ,not the extracted triple. This is the load-bearing design choice that operates before extraction and retains ownership context that no downstream component can recover. A candidate is admitted if it satisfies at least one criterion: (1) it encodes a personaliza- tion relation (preference, habit, ownership, role, or stable affiliation),(2) it encodes a semi-stable user attribute (health, employment, location, or major life event),or (3) it exhibits cross-session reusability ,it would materially improve relevance of a future response. A candidate is rejected if it encodes a transient state, a generic emotional reaction, or vague non-factual content. The following contrasts illustrate why utterance-level evaluation is neces- sary ,each pair produces identical extracted triples: [REJECT] “I drove a Jeep for the road trip that I rented ;” transient,context lost at extraction [ADMIT] “I’ve had my Jeep Wrangler for three years.” stable ownership [REJECT] “I might switch to Android next year.” hedged,no commitment [ADMIT] “I’m switching to Android next month.” defini- tive,stable preference The admission policy outputs a structured JSON object containing the admit/reject decision, entity based ontology mapping, scope classification, and enrichment flag. Every decision is logged and au- ditable. Multi-intent decomposition. A single admitted utterance frequently carries more than one storable proposition, and committing it as one entangled string degrades both retrieval precision and later conflict resolution. Formation therefore decom- poses an admitted exchange into one or more atomic facts, each evaluated and categorized in- dependently. For example, “gift ideas for my son” yields both a persona fact (the user has a son) and a procedural fact (the user searched for gift ideas); “find a hotel with an onsen, and remember I’m veg- etarian” yields a scoped preference and a durable persona attribute. Decomposition is performed at formation, while the utterance is intact, because the boundaries between propositions,and the reference resolution each requires, are themselves part of the context that extraction would otherwise discard. Formation categories. Each atomic fact is as- signed one or more of six categories that determine how the backend should treat it:persona(iden- Figure 1: The Gated Memory pipeline. Every candidate fact passes through an utterance-level admission gate. Admitted facts are scope-classified and conditionally enriched before being passed to the memory backend. tity, relationships, durable attributes),preference (likes, dislikes, dietary and interaction constraints), event(experienced episodes, carrying temporal and participant metadata),procedural(device- or application-level actions the user performed), user_rule(explicit instructions on assistant be- havior), anderror_learning(corrections to prior assistant understanding). Categories are not mutu- ally exclusive: a fact that aids retrieval under more than one is multi-labeled rather than forced into a single class. The categorization is produced jointly with admission so that it, too, is conditioned on the full utterance rather than on the extracted triple. Provenance. Every fact carries a provenance la- bel distinguishing a directly stated attribute, one the user asserts as the point of the utterance (“I have a son”), from an inferred one, extracted from a reference made in passing (“gift ideas for my son” implies, but does not assert, that the user has a son). This distinction, like ownership versus rental, exists only in the original utterance and is unrecoverable downstream. It governs how confidently the back- end may treat a fact: directly stated attributes are candidates for durable semantic storage, whereas inferred attributes are retained provisionally, pend- ing corroboration. Under explicit user instruction to remember, all resulting facts are marked directly stated. Grounding. Formation is constrained to assert only what the evaluated window supports. The pol- icy may not introduce a specific entity or relation absent from the current exchange and its reference context; an unbound reference is resolved to the most specific accurate antecedent, or, failing that, Table 1: Entity scope taxonomy and enrichment policy with privacy constraints. ScopeCoversEnrichment policy USER_PIIHealth,identity,fi- nances, contacts No enrichment. INTERNALDevices, apps, platform data Internal metadata only,no ex- ternal calls EXTERNALProducts, places, public entities External grounding when se- mantically necessary GENERICLLM-resolvable factsParametricknowledgeor none to a generic form, rather than to an invented one. For instance, a window containing only “he is seven” with no antecedent naming a child yields the user’s child is seven, not a fabricated son. This grounding constraint converts a silent failure mode of triple extraction,confident assertion of unsupported enti- ties,into a bounded, auditable decision. Formation also preserves the language of the source utterance, so that a fact stated in one language is stored in that language rather than silently translated. Stage 2: Entity Scope Classification.Every ad- mitted fact is assigned to exactly one of four scopes (Table 1), produced jointly with the admission de- cision and privacy constraint classification. Stage 3: Conditional Enrichment.A fact is en- riched only when two conditions are jointly satis- fied: enrichment would improve its semantic utility for future queries, and the assigned scope permits the required knowledge source. ForEXTERNALen- tities, a single stored fact can support a broad class of downstream queries. For example: “Planning a trip with BMW edrive”→ (Type:ElectricVehicle, ChargingRange:“300 miles”) Deictic grounding. Beyond entity-level enrich- ment, formation resolves relative temporal and spa- tial references against the metadata of the current exchange. Expressions such as “next week,” “to- morrow,” or “tonight” are grounded to absolute dates using the utterance timestamp, and place deixis such as “here” or “there” is resolved to a concrete location using the request context, sub- ject to the same scope-based privacy policy that governs external enrichment. Grounding relative references at formation is, like entity enrichment, a one-time cost amortized across every future re- trieval: a fact whose temporal anchor is resolved once remains interpretable indefinitely, whereas a fact storing “next week” verbatim becomes mean- ingless the moment the utterance’s temporal origin is lost. After enrichment, facts are normalized. Volatile session-specific context is stripped and only stable attributes are committed. Normalized, enriched facts are passed to the memory backend, which handles deduplication, conflict resolution, embed- ding, and persistence. Gated Memory governs what reaches the backend, it does not replace it. Re- trieval and generation are entirely unchanged. 4 Experiments 4.1 Dataset We evaluate on LoCoMo-10, a benchmark of ten multi-session dialogues averaging∼300 turns each, with 1,540 questions across four categories: Single- Hop, Multi-Hop, Temporal, and Open Domain. Category 5 (adversarial) is excluded. 4.2 Baseline and experimental design Our baseline is mem0 framework with a fixed con- figuration of Qdrant vector store, an open-source 8B instruction-tuned LLM and a fixed answer gen- eration prompt. The only experimental variable is the presence of the Gated Memory formation layer, directly isolating formation quality from all other system components. Gated Memory is designed as a production formation layer. The baseline thus represents a complete ablation of the formation layer. Disentangling the individual contributions of admission and enrichment is left as future work. Table 2: LLM-judge accuracy (%) on LoCoMo-10. Re- trieval, generation, and evaluation are identical across both conditions. SystemS-HopM-HopO-DomTempOverall memory baseline49.650.048.357.650.6 Gated Memory51.847.949.958.651.9 Rel Error.∆+4.4%−4.2%+3.3%+1.7%+2.6% 4.3 Metric We report LLM-judge accuracy. An independent evaluator scores each answer as correct or incor- rect against ground truth, with lenient matching for paraphrasing and partial answers. 4.4 Admission statistics The gate processed 5,882 utterances across ten conversations, admitting 5,527 (94.0%) and reject- ing 355 (6.0%). Rejections consisted primarily of vague emotional expressions, procedural turns, and transient situational references. 5 Results and Discussion Table 2 reports results. Gated Memory improves overall accuracy by +2.6% relative with the forma- tion layer as the sole experimental variable, directly attributing the gain to formation quality. Single-Hop and Open Domain. The largest gains appear on Single-Hop (+4.4%) and Open Do- main (+3.3%). Single-Hop performance is directly sensitive to retrieval noise: removing low-signal entries reduces competition for top-kslots, improv- ing precision over the matched retrieval pool. Open Domain additionally benefits from conditional en- richment ,an EV entry augmented with enriched context supports a broad class of queries that the raw extraction string cannot address, and the +3.3% gain quantifies this effect. Temporal. The +1.7% gain on Temporal ques- tions is consistent with the admission policy’s pref- erence for temporally stable facts over episodic states. Facts encoding persistent preferences, and semantic relevance are admitted,transient situa- tional references are rejected. Multi-Hop regression. Multi-Hop exhibits a −4.2% regression. Compositional reasoning is sensitive to both precision and recall. The admis- sion policy, optimized for individual fact utility, occasionally prunes interdependent nodes. This is a precision-recall tradeoff inherent to indepen- dent fact evaluation, not an artifact of our specific threshold, and motivates chain-aware admission that models inter-fact dependencies at formation time as the primary direction for future work. Formation quality as a binding constraint. With retrieval, generation, and evaluation held con- stant, a +2.6% overall gain from changing only what enters the memory store especially for a emo- tionally dense dataset demonstrates that formation quality is a binding constraint on downstream task performance. In actual production environments with high volume of irrelevant context this gating framework could be much more beneficial. 6 Related Work Persistent memory is a very critical component for AI systems Timoneda and Vera (2025); He et al. (2026). Long-term memory architectures for conversational AI have focused mostly on post-formation concerns Zhong et al. (2024); Liu et al. (2023). MemGPT Packer et al. (2023) in- troduces hierarchical memory management with OS-inspired paging while mem0 Chhikara et al. (2025) provides production lifecycle management including deduplication and conflict resolution. A-MEM Xu et al. (2025) proposes Zettelkasten- inspired dynamic interconnection while ReadA- gent Lee et al. (2024) addresses long-context lim- its through gist-based compression. Zep Ras- mussen et al. (2025) uses Graphiti, a temporally- aware knowledge graph engine to dynamically synthesizes both unstructured and structured data while maintaining historical relationships. None of this frameworks addresses the formation qual- ity. Prior Works on retrieval-augmented genera- tion Lewis et al. (2020) and knowledge conflict Shi et al. (2023) has also established that retrieval noise degrades generation quality. Our results extend this causal chain: formation noise→re- trieval noise→generation degradation. Classic work on memory decay Tulving (1987); Baddeley (2000) addresses when facts should be removed,we address whether they should even be admitted. The episodic–semantic distinction Tulving (1987) is typically applied post hoc, as a categorization of already-stored content; we instead determine category, provenance, and applicability scope at formation, while the utterance is intact, because the signals that separate a directly stated attribute from one inferred in passing, or an episodic event from a durable persona attribute, are properties of the original utterance rather than of the extracted triple. Prior lifecycle-oriented systems Packer et al. (2023); Chhikara et al. (2025) operate on extracted triples and cannot recover these signals; the ground- ing constraint we impose,asserting only entities present in the evaluated context,further addresses a fabrication failure mode of triple extraction that downstream deduplication, which only detects con- flicts and duplicates, cannot catch. Privacy research at the training and inference level Carlini et al. (2021) motivates but does not address risks intro- duced by enrichment,our entity scope taxonomy and its policy based enrichment fills this gap. 7 Conclusion We introduced Gated Memory, a formation frame- work that addresses an unexamined failure point in long-term conversational memory. By placing an utterance-level admission gate at the moment a fact is first written to storage and a privacy-aware enrichment stage before extraction, this framework preserves contextual signals irreversibly lost under existing approaches for better memory formation. A +2.6% relative improvement on LoCoMo-10 with no changes to retrieval, generation, or eval- uation establishes formation quality as a measur- able constraint on memory performance, though a−4.2% regression on Multi-Hop reasoning re- veals a precision-recall tradeoff that chain-aware, inter-fact-dependent admission policies should re- solve. In production settings, the gate’s filtering of vacuous content directly reduces storage costs, retrieval latency, and hallucination risk at scale. Domain-adaptive thresholds and knowledge-graph grounding remain natural extensions of future work. The richness of what a personalized assistant can do is bounded by the quality of what it chooses to remember. Limitations The admission gate’s effectiveness depends on the quality of the underlying LLM prompt. Adversar- ial or ambiguous utterances may be misclassified at formation time. Unlike downstream deduplica- tion, formation errors are undetectable after the fact is formed. The evaluation is conducted on LoCoMo-10 with atypical emotional density in utterance data. Hence generalization across do- mains, languages, and task-oriented deployments need to be demonstrated empirically. The 6.0% re- jection rate is a conservative lower bound and task- oriented deployments are expected to exhibit sub- stantially higher filtering rates that needs to be ana- lyzed further. The Multi-Hop regression (−4.2%) reveals a precision-recall tradeoff inherent to in- dependent fact evaluation,chain-aware admission that models inter-fact dependencies remains future work. Finally, the enrichment stage introduces ex- ternal knowledge calls for EXTERNAL-scoped facts, adding latency and potential staleness risk for time- sensitive information that needs to be addressed by only including stable facts for enrichment. References Alan Baddeley. 2000. The episodic buffer: a new com- ponent of working memory? Trends in cognitive sciences, 4(11):417–423. Nicholas Carlini, Florian Tramer, Eric Wallace, Matthew Jagielski, Ariel Herbert-Voss, Katherine Lee, Adam Roberts, Tom Brown, Dawn Song, Ulfar Erlingsson, and 1 others. 2021. Extracting training data from large language models. In 30th USENIX security symposium (USENIX Security 21), pages 2633–2650. Prateek Chhikara, Dev Khant, Saket Aryan, Taranjeet Singh, and Deshraj Yadav. 2025. Mem0: Building production-ready ai agents with scalable long-term memory. arXiv preprint arXiv:2504.19413. Zihong He, Weizhe Lin, Hao Zheng, Fan Zhang, Matt W. Jones, Laurence Aitchison, Xuhai Xu, Miao Liu, Hai-Ning Liang, Per Ola Kristensson, and Junxiao Shen. 2026. Human-inspired perspectives: A survey on ai long-term memory. Proceedings of the IEEE, 114(3):484–515. Kuang-Huei Lee, Xinyun Chen, Hiroki Furuta, John Canny, and Ian Fischer. 2024. A human-inspired reading agent with gist memory of very long contexts. arXiv preprint arXiv:2402.09727. Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Hein- rich Küttler, Mike Lewis, Wen-tau Yih, Tim Rock- täschel, and 1 others. 2020. Retrieval-augmented gen- eration for knowledge-intensive nlp tasks. Advances in neural information processing systems, 33:9459– 9474. Lei Liu, Xiaoyan Yang, Yue Shen, Binbin Hu, Zhiqiang Zhang, Jinjie Gu, and Guannan Zhang. 2023. Think- in-memory: Recalling and post-thinking enable llms with long-term memory.arXiv preprint arXiv:2311.08719. Adyasha Maharana, Dong-Ho Lee, Sergey Tulyakov, Mohit Bansal, Francesco Barbieri, and Yuwei Fang. 2024.Evaluating very long-term conversational memory of llm agents. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 13851– 13870. Charles Packer, Sarah Wooders, Kevin Lin, Vivian Fang, Shishir G Patil, Ion Stoica, and Joseph E Gonzalez. 2023. Memgpt: Towards llms as operating systems. arXiv preprint arXiv:2310.08560. Preston Rasmussen, Pavlo Paliychuk, Travis Beauvais, Jack Ryan, and Daniel Chalef. 2025. Zep: a tempo- ral knowledge graph architecture for agent memory. arXiv preprint arXiv:2501.13956. Freda Shi, Xinyun Chen, Kanishka Misra, Nathan Scales, David Dohan, Ed H Chi, Nathanael Schärli, and Denny Zhou. 2023. Large language models can be easily distracted by irrelevant context. In Inter- national Conference on Machine Learning, pages 31210–31227. PMLR. Joan C Timoneda and Sebastián Vallejo Vera. 2025. Memory is all you need: Testing how model memory affects llm performance in annotation tasks. arXiv preprint arXiv:2503.04874. Endel Tulving. 1987. Multiple memory systems and consciousness. Human neurobiology, 6(2):67–80. Wujiang Xu, Zujie Liang, Kai Mei, Hang Gao, Jun- tao Tan, and Yongfeng Zhang. 2025.A-mem: Agentic memory for llm agents. arXiv preprint arXiv:2502.12110. Wanjun Zhong, Lianghong Guo, Qiqi Gao, He Ye, and Yanlin Wang. 2024. Memorybank: Enhancing large language models with long-term memory. In Pro- ceedings of the AAAI Conference on Artificial Intelli- gence, volume 38, pages 19724–19731.