Pith. sign in

REVIEW 5 major objections 6 minor 67 references

A framework that lets memory components be swapped across agent systems reveals which design choices actually matter.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 07:30 UTC pith:GE6RHHTL

load-bearing objection A useful framework paper whose empirical claims are thinner than the architecture; worth refereeing, but the demonstrations don't yet secure the 'systematic isolation' promise. the 5 major comments →

arxiv 2607.21404 v1 pith:GE6RHHTL submitted 2026-07-23 cs.CL

MemTools: A Unified Research Framework for Interoperable Agent Memory

Chengfeng Zhao , Jinhui Chen , Sirui Liang , Shizhu He , Yequan Wang , Jun Zhao , Kang Liu This is my paper
classification cs.CL
keywords agent memorymemory lifecycledeclarative data contractsinteroperabilityevaluation protocolsheterogeneous memoryLLM agentsmodular pipelines
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Agent memory research is fragmented: systems fuse memory formation, storage, retrieval, and evolution with each other and with specific benchmarks, so differences in performance cannot be traced to any one design decision. MemTools attacks this by standardizing the memory lifecycle with declarative data contracts, separating evaluation protocols from datasets, and unifying symbolic, neural, and multimodal memory under a single interface. Three experiments show the payoff: a hybrid pipeline built from components of two different systems beats the native design; the same pipeline run under batch versus streaming protocols drops from 40.30 to 33.58, isolating the timing of memory formation as the only cause; and coordinating two memory representations outperforms either alone. If the framework works as claimed, it gives researchers a practical way to vary one memory-design variable at a time and measure its true effect.

Core claim

The paper's central claim is that decoupling memory systems from their deployment environments — through standardized stage interfaces, pluggable evaluation protocols, and a shared coordination layer — makes agent-memory research systematic rather than anecdotal. The empirical evidence is that cross-system component assemblies can match or exceed native designs (43.28 vs 40.30 on ALFWorld), that the performance gap between batch and streaming memory formation (40.30 vs 33.58) is entirely attributable to formation timing because all other components and data remain fixed, and that coordinated heterogeneous memory provides complementary gains (35.3 vs 30.1/31.2 on Mem-Gallery; 51.5 vs 48.5/32.

What carries the argument

The load-bearing mechanism is the declarative data contract: every component declares its requires_keys and provides_keys, and a matching engine validates a pipeline by checking that each downstream component's required fields are a subset of what upstream components provide. This machinery converts component assembly into a checkable declaration, automatically enumerates valid pipeline combinations, and applies the same contract to benchmark–protocol matching. A second key piece is the orthogonal split between benchmark datasets (data providers) and evaluation protocols (execution logic), which allows the same pipeline to run under different memory-formation timings. The MultiSystem coordin

Load-bearing premise

The central premise is that structural field matching (requires_keys ⊆ provides_keys) is sufficient for two components to work correctly together, although the authors concede that subtle behavioral differences between systems can cause misalignment even when structures match.

What would settle it

A concrete test: assemble a pipeline whose components declare identical field names but internally expect different formats (e.g., one component reads 'text_content' as raw text, another expects pre-tokenized input in the same field); if the framework accepts the pipeline and produces degraded or invalid outputs without error, structural compatibility is demonstrated to be insufficient. A second test: re-run the Batch-vs-Stream protocol comparison on two other benchmark datasets and check whether the 40.30→33.58 drop is consistently attributable to formation timing.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Researchers can run the same memory components under multiple evaluation protocols to separate intrinsic memory quality from when and how memory is written.
  • Hybrid pipelines assembled from different systems become first-class experimental conditions rather than engineering one-offs, enabling controlled cross-system comparisons.
  • Adding a new memory representation (symbolic, neural, or multimodal) reduces to building a pipeline that satisfies the lifecycle contract, rather than integrating with each host system.
  • Reported performance claims in agent-memory papers can be re-tested by swapping protocols or components, making results more reproducible and easier to attribute.
  • The framework's compatibility checks cut developer effort — 60.1% fewer lines of code and 85.4% fewer custom functions in the reported user study — lowering the cost of ablation studies.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Field-name overlap is a proxy, not proof, of true compatibility: two components that both declare a 'text' field may internally expect different encodings, and the framework's structural check would let them assemble while silently misbehaving — the authors acknowledge this; a stress test with near-mismatched contracts would sharpen the boundary.
  • The Batch-to-Stream performance drop suggests continuous incremental memory formation can be actively harmful, not merely neutral; this hypothesis is directly testable on additional benchmarks and could motivate memory systems that throttle or batch formation.
  • The coordinated-memory gains may come from complementary recall rather than from the coordination layer itself; an ablation that feeds the same two pipelines into the model sequentially without synchronization would separate those explanations.
  • The three demonstrations use specific base systems; reproducing them with a different set of adapters (e.g., other backends and retrieval modules) would test whether the interoperability claims generalize beyond the reported configuration.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

5 major / 6 minor

Summary. MemTools is an interoperability framework for agent-memory research. It refactors the memory lifecycle into formation, retrieval, evolution, and utilization stages connected by declarative data contracts (requires_keys/provides_keys), decouples benchmark datasets from evaluation protocols, and provides a MultiSystem coordination layer for symbolic, neural, and multimodal memory. The paper reports three empirical demonstrations: (1) a cross-system hybrid (AWM formation with A-Mem backend/retrieval) outperforming a native AWM pipeline on ALFWorld (Table 2); (2) a Batch-to-Stream performance drop attributed entirely to the timing of memory formation (Section 4.2); and (3) coordinated heterogeneous memory improving over isolated symbolic/neural/multimodal pipelines (Figure 6). A small developer-effort user study in Appendix D claims 60.1% LOC reduction and 85.4% reduction in custom definitions. The paper concludes that MemTools enables systematic isolation and analysis of memory design variables.

Significance. If the empirical claims were fully supported, MemTools would be a useful infrastructure contribution: the architecture is coherent, the API in Figure 3 is simple, the dataset/protocol decoupling is a sensible design principle, the code is open-source, and the Batch-to-Stream result honestly replicates an externally reported effect (Zhang et al., 2026). However, the current evidence is not strong enough to establish the central claim. The headline numbers are single-run point estimates with no variance, significance tests, or seed information; the structural compatibility check is explicitly conceded to be only structural; and the developer-effort study uses a small, poorly defined sample. Thus the significance is conditional: the framework has real potential, but the paper's empirical backbone needs substantial reinforcement before the 'systematic isolation and analysis' claim can be accepted.

major comments (5)
  1. [§4.1, Table 2] Rows 1 and 2 compare a native AWM pipeline (40.30) with a cross-system hybrid (43.28) using what appear to be single-run point estimates. No number of runs, seeds, temperature values, episode counts, or significance tests are reported. A 2.98-point difference with no variance information cannot support the conclusion that 'cross-system component assemblies can match or exceed native designs.' Please report repeated runs with confidence intervals and a significance test, and specify the full evaluation configuration (LLM backbone, decoding parameters, number of tasks, seeds).
  2. [§4.2, Table 2] The claim that the Batch-to-Stream drop (40.30 vs 33.58) 'is entirely attributable to the timing of memory formation' is stronger than the evidence supports. It assumes both that the protocol abstraction changes only the timing of memory formation and that the AWM components behave identically inside and outside their native host. Neither is demonstrated. The single-run comparison also gives no estimate of noise. Please provide a protocol-ablation study that verifies all memory operations other than timing remain unchanged, and add statistical support for the observed drop.
  3. [§4.3, Figure 6] The coordinated-memory gains (35.3 vs 30.1/31.2 on Mem-Gallery; 51.5 vs 48.5/32.1 on ALFWorld) are reported as single point estimates with no variance or significance testing. The conclusion that distinct memory representations 'provide complementary capabilities' is not statistically established. Additionally, the coordination layer may give the system access to strictly more information than any isolated subsystem; without controlling for this or reporting how retrieved content is merged, the measured gains could reflect additional context rather than true representational complementarity.
  4. [§3.1, Appendix A, Limitations] The matching engine verifies only that upstream provides_keys is a superset of downstream requires_keys. The Limitations section explicitly concedes that 'the automatic matching engine only verifies the structural compatibility of data fields. Subtle system differences may cause behavioral misalignments between components.' This limitation is load-bearing because every empirical demonstration assumes assembled hybrid or coordinated components behave as their native counterparts would. A field named text_content may have different serialization, chunking, or ranking behavior under a foreign backend, and the adapter could introduce write-timing or tie-breaking differences. Please add behavioral validation: for each adapter, unit tests comparing component outputs on identical inputs inside and outside the native host, or a reported quantification of adapter-induced behavioral misalignment.
  5. [Appendix D, Table 3] The developer-effort study reports n=3, an undefined 'Native' condition, and a very large standard deviation (±94 LOC) relative to the mean (135.5). The 60.1% and 85.4% reductions are therefore not interpretable. Please define the Native baseline concretely (which library, which pipeline, what code is counted), report per-annotator results, describe the exact task, and use an appropriate statistical comparison. If this study is meant to support the framework's practical utility, it needs the same rigor as the main experiments.
minor comments (6)
  1. [§4.1] The sentence 'This performance variance highlights...' uses 'variance' to mean difference between two point estimates; no variance is actually reported. Consider rewording to 'performance difference'.
  2. [Figure 4] The percentages in the incompatibility-cause pie chart (37.0%, 20.1%, 23.9%, 19.0%) are given without denominators or a description of how failures were classified. Please state N and the classification procedure.
  3. [Table 2, Figure 6] The numeric values in Table 2 and Figure 6 are called 'Success Rate' but no units are specified. State explicitly whether these are percentages, scores in [0,100], or raw counts, and label the Figure 6 y-axes consistently.
  4. [Table 1] The legend uses three visually similar symbols (●, ●, ●) for full, partial, and no support. This is difficult to read in print; please use clearly distinct markers or add text labels.
  5. [Reproducibility] Add a reproducibility statement listing the exact LLM versions, sampling parameters, random seeds, and number of evaluation tasks for all experiments. This is currently absent and is needed for any comparative claim.
  6. [Appendix D] The term 'LOC' in Table 3 conflicts with the prose 'effective lines of code.' Clarify the counting rule and, if possible, provide the exact code for both the Native and MemTools conditions.

Circularity Check

0 steps flagged

No significant circularity; the empirical claims are measurements and the key batch-to-stream effect is explicitly aligned with external prior work.

full rationale

The paper's central empirical claims (Table 2, Figure 4, Figure 6) are direct measurements of assembled pipelines, not predictions derived from fitted parameters or from the framework's own definitions. The compatibility criterion (provides_keys superset of requires_keys, Appendix A) is explicitly structural, and the Limitations section concedes that 'its automatic matching engine only verifies the structural compatibility of data fields. Subtle system differences may cause behavioral misalignments between components.' That is a validity limitation, not a circular reduction: the paper does not claim the matching engine guarantees behavioral equivalence, and the empirical numbers would be meaningful measurements even if the limitation weakens their interpretation. The only self-citations (e.g., NPM, Zhao et al. 2026; DyPRAG, Tan et al. 2025) are used as integrated external components; their behavior is not derived from MemTools, and no uniqueness or forced-choice argument rests on them. Section 4.2's attribution of the batch-to-stream drop to formation timing is a controlled single-variable comparison, and the paper explicitly aligns the result with the external finding of Zhang et al. 2026 rather than presenting it as a novel first-principles derivation. No equation or result reduces to its own input.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 2 invented entities

The framework's central claims rest on design choices and assumptions about the memory domain rather than on fitted math. The headline numbers depend on unstated evaluation configurations (LLM, seeds, episode counts); the lifecycle ontology (formation/retrieval/evolution/utilization/backend) is an assumption about how agent memory decomposes; compatibility is defined as structural key matching, which the authors themselves concede is weaker than behavioral compatibility; and the Batch-vs-Stream attribution assumes protocol changes alter nothing but formation timing. No physical entities are postulated; the framework abstractions (data contracts, coordination layer) are software constructs whose evidence is confined to this paper.

free parameters (2)
  • Evaluation configuration (LLM backbone, seeds, temperature, episode counts, run counts) = unreported (code shows 'model_name=xxx'; max_test_tasks=10)
    Headline results in Table 2 and Figure 6 depend on hand-chosen test settings that are never specified, so the single-run deltas cannot be interpreted or reproduced.
  • AWM formation max_rules cap = 20 (usage example)
    Rule-extraction limit for the AWM formation module shown in Figure 3; controls how much of the trajectory is distilled and therefore affects the hybrid pipeline score in Table 2; not reported as an experimental variable.
axioms (4)
  • domain assumption Agent memory can be canonically decomposed into formation, retrieval, evolution, utilization, and backend stages
    Section 3.1 structures the entire framework around this lifecycle ontology; if the right decomposition is different (e.g., memory as continuous state rather than discrete entries), the standardization encodes the wrong seams.
  • domain assumption Structurally matching provides_keys ⊇ requires_keys is a sufficient condition for pipeline interoperability
    Appendix A defines compatibility this way; the authors' Limitations paragraph concedes it only checks structure and can miss behavioral misalignment.
  • domain assumption The Batch-vs-Stream performance gap is caused entirely by memory-formation timing
    Section 4.2 asserts the 40.30-to-33.58 drop is 'entirely attributable to the timing of memory formation'; this requires that protocol changes alter no other variable (context length, retrieval timing, decision points).
  • domain assumption Prior benchmarks (ALFWorld, Mem-Gallery) yield stable single-run estimates adequate for comparing configurations
    Tables 2 and Figure 6 report single numbers with no seed variance; the framework's demonstrations assume one run per configuration is enough to support conclusions like 'coordinated > isolated'.
invented entities (2)
  • Declarative data contract (requires_keys/provides_keys vocabulary) no independent evidence
    purpose: Canonical inter-component interface enabling automatic compatibility checking and pipeline enumeration
    A software abstraction with no evidence beyond this paper's own demonstrations; its sufficiency claim is contradicted in part by the authors' own Limitations note.
  • MultiSystem coordination layer no independent evidence
    purpose: Unified interface running symbolic, neural, and multimodal pipelines in parallel with synchronized search and chained utilization
    The claimed complementarity gains (Figure 6) are its only evidence, but with no variance estimates and a possible confound (coordinated systems retrieve from more sources simultaneously), the entity's benefit is not independently established.

pith-pipeline@v1.3.0-alltime-deepseek · 11668 in / 17858 out tokens · 166447 ms · 2026-08-01T07:30:56.647877+00:00 · methodology

0 comments
read the original abstract

While memory systems are essential for agent architectures, pervasive architectural fragmentation restricts systematic research. Existing implementations typically couple different stages of the memory lifecycle, entangle evaluation logic with specific datasets, and provide limited support for the management of heterogeneous memory types. We introduce MemTools, an interoperability research framework that decouples memory system components from their underlying deployment environments. MemTools standardizes the memory lifecycle through declarative data contracts, enabling the interchangeable assembly of components across different systems. It orthogonally separates benchmark datasets from execution protocols to facilitate controlled assessments. Furthermore, MemTools provides a unified computational interface for coordinating symbolic, neural, and multimodal memory representations within a shared runtime. Empirical evaluations on cross-system component integration, evaluation protocol reconfiguration, and heterogeneous memory coordination demonstrate that MemTools enables systematic isolation and analysis of memory design variables. These findings suggest that MemTools provides a practical and extensible infrastructure for advancing principled research on agent memory.

Figures

Figures reproduced from arXiv: 2607.21404 by Chengfeng Zhao, Jinhui Chen, Jun Zhao, Kang Liu, Shizhu He, Sirui Liang, Yequan Wang.

Figure 1
Figure 1. Figure 1: The MemTools architecture. Heterogeneous memory coordination unifies symbolic, neural, and multi [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Data flows of different evaluation protocols. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Distribution of pipeline incompatibility causes [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: Performance comparison of isolated mem￾ory representations versus coordinated multi-system pipelines. The left axis reports F1 scores on Mem￾Gallery for symbolic and multimodal coordination, while the right axis reports success rates on ALFWorld for symbolic and neural coordination. demand symbolic, neural, and multimodal mem￾ory simultaneously [PITH_FULL_IMAGE:figures/full_fig_p006_6.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

67 extracted references · 4 canonical work pages

  1. [1]

    2025 , eprint=

    Memory in the Age of AI Agents , author=. 2025 , eprint=

    2025

  2. [2]

    2026 , eprint=

    Rethinking Memory Mechanisms of Foundation Agents in the Second Half: A Survey , author=. 2026 , eprint=

    2026

  3. [3]

    2026 , eprint=

    Externalization in LLM Agents: A Unified Review of Memory, Skills, Protocols and Harness Engineering , author=. 2026 , eprint=

    2026

  4. [4]

    2024 , eprint=

    MemGPT: Towards LLMs as Operating Systems , author=. 2024 , eprint=

    2024

  5. [5]

    2025 , eprint=

    Zep: A Temporal Knowledge Graph Architecture for Agent Memory , author=. 2025 , eprint=

    2025

  6. [6]

    2025 , eprint=

    MemOS: A Memory OS for AI System , author=. 2025 , eprint=

    2025

  7. [7]

    2025 , eprint=

    Mem0: Building Production-Ready AI Agents with Scalable Long-Term Memory , author=. 2025 , eprint=

    2025

  8. [8]

    2025 , eprint=

    MIRIX: Multi-Agent Memory System for LLM-Based Agents , author=. 2025 , eprint=

    2025

  9. [9]

    2026 , eprint=

    Omni-SimpleMem: Autoresearch-Guided Discovery of Lifelong Multimodal Agent Memory , author=. 2026 , eprint=

    2026

  10. [10]

    The Fourteenth International Conference on Learning Representations , year=

    Seeing, Listening, Remembering, and Reasoning: A Multimodal Agent with Long-Term Memory , author=. The Fourteenth International Conference on Learning Representations , year=

  11. [11]

    Know- MRI : A Knowledge Mechanisms Revealer & Interpreter for Large Language Models

    Liu, Jiaxiang and Xing, Boxuan and Yuan, Chenhao and Zhang, Chenxiang and Wu, Di and Huang, Xiusheng and Yu, Haida and Lang, Chuhan and Cao, Pengfei and Zhao, Jun and Liu, Kang. Know- MRI : A Knowledge Mechanisms Revealer & Interpreter for Large Language Models. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volum...

    work page doi:10.18653/v1/2025.acl-demo.20 2025

  12. [12]

    Proceedings of the 42nd International Conference on Machine Learning , articleno =

    Wang, Zora Zhiruo and Mao, Jiayuan and Fried, Daniel and Neubig, Graham , title =. Proceedings of the 42nd International Conference on Machine Learning , articleno =. 2025 , publisher =

    2025

  13. [13]

    2026 , eprint=

    SimpleMem: Efficient Lifelong Memory for LLM Agents , author=. 2026 , eprint=

    2026

  14. [14]

    Zhao, Andrew and Huang, Daniel and Xu, Quentin and Lin, Matthieu and Liu, Yong-Jin and Huang, Gao , title =. Proceedings of the Thirty-Eighth AAAI Conference on Artificial Intelligence and Thirty-Sixth Conference on Innovative Applications of Artificial Intelligence and Fourteenth Symposium on Educational Advances in Artificial Intelligence , articleno =....

    doi:10.1609/aaai.v38i17.29936 2024

  15. [15]

    arXiv preprint arXiv:2605.13941 , year=

    EvolveMem: Self-Evolving Memory Architecture via AutoResearch for LLM Agents , author=. arXiv preprint arXiv:2605.13941 , year=

    Pith/arXiv arXiv

  16. [16]

    2026 , eprint=

    Evo-Memory: Benchmarking LLM Agent Test-time Learning with Self-Evolving Memory , author=. 2026 , eprint=

    2026

  17. [17]

    E asy E dit: An Easy-to-use Knowledge Editing Framework for Large Language Models

    Wang, Peng and Zhang, Ningyu and Tian, Bozhong and Xi, Zekun and Yao, Yunzhi and Xu, Ziwen and Wang, Mengru and Mao, Shengyu and Wang, Xiaohan and Cheng, Siyuan and Liu, Kangwei and Ni, Yuansheng and Zheng, Guozhou and Chen, Huajun. E asy E dit: An Easy-to-use Knowledge Editing Framework for Large Language Models. Proceedings of the 62nd Annual Meeting of...

    doi:10.18653/v1/2024.acl-demos.9 2024

  18. [18]

    Evaluating Very Long-Term Conversational Memory of LLM Agents

    Maharana, Adyasha and Lee, Dong-Ho and Tulyakov, Sergey and Bansal, Mohit and Barbieri, Francesco and Fang, Yuwei. 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). 2024. doi:10.18653/v1/2024.acl-long.747

    doi:10.18653/v1/2024.acl-long.747 2024

  19. [19]

    2025 , eprint=

    LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory , author=. 2025 , eprint=

    2025

  20. [20]

    2021 , url =

    Mohit Shridhar and Xingdi Yuan and Marc-Alexandre C\^ot\'e and Yonatan Bisk and Adam Trischler and Matthew Hausknecht , booktitle =. 2021 , url =

    2021

  21. [21]

    Proceedings of the AAAI Conference on Artificial Intelligence , author=

    MemoryBank: Enhancing Large Language Models with Long-Term Memory , volume=. Proceedings of the AAAI Conference on Artificial Intelligence , author=. 2024 , month=. doi:10.1609/aaai.v38i17.29946 , number=

    doi:10.1609/aaai.v38i17.29946 2024

  22. [22]

    arXiv preprint arXiv:2307.13854 , url=

    WebArena: A Realistic Web Environment for Building Autonomous Agents , author=. arXiv preprint arXiv:2307.13854 , url=

    Pith/arXiv arXiv

  23. [23]

    2025 , eprint=

    Advances and Challenges in Foundation Agents: From Brain-Inspired Intelligence to Evolutionary, Collaborative, and Safe Systems , author=. 2025 , eprint=

    2025

  24. [24]

    2023 , journal =

    Voyager: An Open-Ended Embodied Agent with Large Language Models , author =. 2023 , journal =

    2023

  25. [25]

    Frontiers of Computer Science , volume=

    A survey on large language model based autonomous agents , author=. Frontiers of Computer Science , volume=. 2024 , publisher=

    2024

  26. [26]

    arXiv preprint arXiv:2401.03428 , year=

    Exploring large language model based intelligent agents: Definitions, methods, and prospects , author=. arXiv preprint arXiv:2401.03428 , year=

    Pith/arXiv arXiv

  27. [27]

    The Twelfth International Conference on Learning Representations , year=

    Synapse: Trajectory-as-Exemplar Prompting with Memory for Computer Control , author=. The Twelfth International Conference on Learning Representations , year=

  28. [28]

    2025 , eprint=

    ReasoningBank: Scaling Agent Self-Evolving with Reasoning Memory , author=. 2025 , eprint=

    2025

  29. [29]

    Proceedings of the 41st International Conference on Machine Learning , articleno =

    Wang, Yu and Gao, Yifan and Chen, Xiusi and Jiang, Haoming and Li, Shiyang and Yang, Jingfeng and Yin, Qingyu and Li, Zheng and Li, Xian and Yin, Bing and Shang, Jingbo and McAuley, Julian , title =. Proceedings of the 41st International Conference on Machine Learning , articleno =. 2024 , publisher =

    2024

  30. [30]

    M+: Extending Memory

    Yu Wang and Dmitry Krotov and Yuanzhe Hu and Yifan Gao and Wangchunshu Zhou and Julian McAuley and Dan Gutfreund and Rogerio Feris and Zexue He , booktitle=. M+: Extending Memory. 2025 , url=

    2025

  31. [31]

    2025 , eprint=

    MemGen: Weaving Generative Latent Memory for Self-Evolving Agents , author=. 2025 , eprint=

    2025

  32. [32]

    2025 , eprint=

    VisMem: Latent Vision Memory Unlocks Potential of Vision-Language Models , author=. 2025 , eprint=

    2025

  33. [33]

    Thirty-seventh Conference on Neural Information Processing Systems , year=

    Reflexion: language agents with verbal reinforcement learning , author=. Thirty-seventh Conference on Neural Information Processing Systems , year=

  34. [34]

    2025 , eprint=

    The Illusion of Diminishing Returns: Measuring Long Horizon Execution in LLMs , author=. 2025 , eprint=

    2025

  35. [35]

    Trends in cognitive sciences , volume=

    Dissociating language and thought in large language models , author=. Trends in cognitive sciences , volume=. 2024 , publisher=

    2024

  36. [36]

    2025 , eprint=

    Agent Learning via Early Experience , author=. 2025 , eprint=

    2025

  37. [37]

    The Thirteenth International Conference on Learning Representations , year=

    Self-Updatable Large Language Models by Integrating Context into Model Parameters , author=. The Thirteenth International Conference on Learning Representations , year=

  38. [38]

    Advances in Neural Information Processing Systems , volume=

    Webshop: Towards scalable real-world web interaction with grounded language agents , author=. Advances in Neural Information Processing Systems , volume=

  39. [39]

    arXiv preprint arXiv:2505.09388 , year=

    Qwen3 Technical Report , author=. arXiv preprint arXiv:2505.09388 , year=

    Pith/arXiv arXiv

  40. [40]

    Bernal Jimenez Gutierrez and Yiheng Shu and Yu Gu and Michihiro Yasunaga and Yu Su , booktitle=. Hippo. 2024 , url=

    2024

  41. [41]

    , title =

    Park, Joon Sung and O'Brien, Joseph and Cai, Carrie Jun and Morris, Meredith Ringel and Liang, Percy and Bernstein, Michael S. , title =. Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology , articleno =. 2023 , isbn =. doi:10.1145/3586183.3606763 , abstract =

    arXiv 2023

  42. [42]

    2025 , eprint=

    LEGOMem: Modular Procedural Memory for Multi-agent LLM Systems for Workflow Automation , author=. 2025 , eprint=

    2025

  43. [43]

    ACM Trans

    Zhang, Zeyu and Dai, Quanyu and Bo, Xiaohe and Ma, Chen and Li, Rui and Chen, Xu and Zhu, Jieming and Dong, Zhenhua and Wen, Ji-Rong , title =. ACM Trans. Inf. Syst. , month = sep, articleno =. 2025 , issue_date =. doi:10.1145/3748302 , abstract =

    doi:10.1145/3748302 2025

  44. [44]

    Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing , pages=

    Scienceworld: Is your agent smarter than a 5th grader? , author=. Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing , pages=

    2022

  45. [45]

    Maxime Chevalier-Boisvert and Dzmitry Bahdanau and Salem Lahlou and Lucas Willems and Chitwan Saharia and Thien Huu Nguyen and Yoshua Bengio , booktitle=. Baby. 2019 , url=

    2019

  46. [46]

    Transactions on Machine Learning Research , issn=

    Personalization of Large Language Models: A Survey , author=. Transactions on Machine Learning Research , issn=. 2025 , url=

    2025

  47. [47]

    A-Mem: Agentic Memory for

    Wujiang Xu and Zujie Liang and Kai Mei and Hang Gao and Juntao Tan and Yongfeng Zhang , booktitle=. A-Mem: Agentic Memory for. 2026 , url=

    2026

  48. [48]

    2025 , eprint=

    Echo: A Large Language Model with Temporal Episodic Memory , author=. 2025 , eprint=

    2025

  49. [49]

    Bernal Jim. From. Forty-second International Conference on Machine Learning , year=

  50. [50]

    2024 , url=

    Carlos E Jimenez and John Yang and Alexander Wettig and Shunyu Yao and Kexin Pei and Ofir Press and Karthik R Narasimhan , booktitle=. 2024 , url=

    2024

  51. [51]

    2026 , eprint=

    LatentSkill: From In-Context Textual Skills to In-Weight Latent Skills for LLM Agents , author=. 2026 , eprint=

    2026

  52. [52]

    2026 , eprint=

    MemFactory: Unified Inference & Training Framework for Agent Memory , author=. 2026 , eprint=

    2026

  53. [53]

    arXiv preprint arXiv:2505.02099 , year=

    MemEngine: A Unified and Modular Library for Developing Advanced Memory of LLM-based Agents , author=. arXiv preprint arXiv:2505.02099 , year=

    Pith/arXiv arXiv

  54. [54]

    2026 , eprint=

    Remember Me, Refine Me: A Dynamic Procedural Memory Framework for Experience-Driven Agent Evolution , author=. 2026 , eprint=

    2026

  55. [55]

    2026 , eprint=

    Useful Memories Become Faulty When Continuously Updated by LLMs , author=. 2026 , eprint=

    2026

  56. [56]

    Mem-Gallery: Benchmarking Multimodal Long-Term Conversational Memory for MLLM Agents

    Bei, Yuanchen and Wei, Tianxin and Ning, Xuying and Zhao, Yanjun and Liu, Zhining and Lin, Xiao and Zhu, Yada and Hamann, Hendrik and He, Jingrui and Tong, Hanghang. Mem-Gallery: Benchmarking Multimodal Long-Term Conversational Memory for MLLM Agents. Proceedings of the 64th Annual Meeting of the A ssociation for C omputational L inguistics (Volume 1: Lon...

    work page doi:10.18653/v1/2026.acl-long.1892 2026

  57. [57]

    2026 , eprint=

    Neural Procedural Memory: Empowering LLM Agents with Implicit Activation Steering , author=. 2026 , eprint=

    2026

  58. [58]

    From Storage to Experience: A Survey on the Evolution of LLM Agent Memory Mechanisms

    Luo, Jinghao and Tian, Yuchen and Cao, Chuxue and Luo, Ziyang and Lin, Hongzhan and Li, Kaixin and Kong, Chuyi and Yang, Ruichao and Ma, Jing. From Storage to Experience: A Survey on the Evolution of LLM Agent Memory Mechanisms. Findings of the A ssociation for C omputational L inguistics: ACL 2026. 2026. doi:10.18653/v1/2026.findings-acl.2069

    work page doi:10.18653/v1/2026.findings-acl.2069 2026

  59. [59]

    2026 , publisher=

    Llm agent memory: A survey from a unified representation--management perspective , author=. 2026 , publisher=

    2026

  60. [60]

    2026 , eprint=

    EvoMemBench: Benchmarking Agent Memory from a Self-Evolving Perspective , author=. 2026 , eprint=

    2026

  61. [61]

    Beyond Goldfish Memory: Long-Term Open-Domain Conversation

    Xu, Jing and Szlam, Arthur and Weston, Jason. Beyond Goldfish Memory: Long-Term Open-Domain Conversation. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2022. doi:10.18653/v1/2022.acl-long.356

    doi:10.18653/v1/2022.acl-long.356 2022

  62. [62]

    MQ u AKE : Assessing Knowledge Editing in Language Models via Multi-Hop Questions

    Zhong, Zexuan and Wu, Zhengxuan and Manning, Christopher and Potts, Christopher and Chen, Danqi. MQ u AKE : Assessing Knowledge Editing in Language Models via Multi-Hop Questions. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. 2023. doi:10.18653/v1/2023.emnlp-main.971

    doi:10.18653/v1/2023.emnlp-main.971 2023

  63. [63]

    2020 , url =

    Mohit Shridhar and Jesse Thomason and Daniel Gordon and Yonatan Bisk and Winson Han and Roozbeh Mottaghi and Luke Zettlemoyer and Dieter Fox , booktitle =. 2020 , url =

    2020

  64. [64]

    Proceedings of the 38th International Conference on Neural Information Processing Systems , articleno =

    Kuratov, Yuri and Bulatov, Aydar and Anokhin, Petr and Rodkin, Ivan and Sorokin, Dmitry and Sorokin, Artyom and Burtsev, Mikhail , title =. Proceedings of the 38th International Conference on Neural Information Processing Systems , articleno =. 2024 , isbn =

    2024

  65. [65]

    2026 , eprint=

    LightMem: Lightweight and Efficient Memory-Augmented Generation , author=. 2026 , eprint=

    2026

  66. [66]

    S truct M em: Structured Memory for Long-Horizon Behavior in LLM s

    Xu, Buqiang and Chen, Yijun and Fang, Jizhan and Zhong, Ruobin and Yao, Yunzhi and Zhu, Yuqi and Du, Lun and Deng, Shumin. S truct M em: Structured Memory for Long-Horizon Behavior in LLM s. Proceedings of the 64th Annual Meeting of the A ssociation for C omputational L inguistics (Volume 2: Short Papers). 2026. doi:10.18653/v1/2026.acl-short.12

    work page doi:10.18653/v1/2026.acl-short.12 2026

  67. [67]

    2025 , eprint=

    Dynamic Parametric Retrieval Augmented Generation for Test-time Knowledge Enhancement , author=. 2025 , eprint=

    2025