Pith. sign in

REVIEW 3 major objections 4 minor 40 references

MemTxn argues that agent memory needs a transaction boundary: gate each write against its source, pick the visible version by a declared chronology rule, and restore the complete state after a fault — and shows the gate is exactly correct o

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 00:31 UTC pith:DSTN4ADI

load-bearing objection A real, honestly scoped systems contribution that deserves peer review, but the headline admission-gate claim outruns the evidence until natural-update evaluation or code is released. the 3 major comments →

arxiv 2607.27834 v1 pith:DSTN4ADI submitted 2026-07-30 cs.AI cs.CL

MemTxn: A Transaction Boundary for Source-Supported Updates and Complete-State Recovery in Agent Memory

Hanshuai Cui , Zhiqing Tang , Zhi Yao , Fanshuai Meng , Qianli Ma , Weijia Jia This is my paper
classification cs.AI cs.CL
keywords agent memorytransaction boundarysource-supported updateversion resolutionsnapshot recoveryordered subsequencepersistent memorygovernance layer
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.

MemTxn sets out to make writable agent memory as reliable as a database transaction. The paper's claim is three-fold: an update should be admitted only when the proposed value is an ordered, verbatim subsequence of the cited evidence; when separate sources conflict, the version with the latest declared chronology should be visible; and after a durable fault the system should restore the complete previously active state without knowing which physical writes were lost. The paper builds this boundary outside the answer model so that rejection, fallback, and rollback remain observable. If correct, this gives a label-free, verifier-free governance layer that blocks subtle extraction errors before they become permanent, and makes crash recovery complete rather than partial. The evaluation supports the mechanism-level claims under the declared lexical and chronological contracts, with the paper explicitly scoping to those contracts rather than to semantic truth.

Core claim

The central discovery is that the hard part of reliable agent memory is not storage atomicity but the semantic contract around it. MemTxn encodes that contract as three deterministic, answer-independent rules. Ordered PatchTest, defined by Support_ord(q)=F(q)∧Q(e,D_s)∧[T(v)⪯T(e)], requires every token of the proposed value to occur within the cited evidence span in source order; on a held-out audit it accepted all 60 supported originals and rejected all 179 hard negatives that preserved word overlap but changed order or polarity. The Temporal Resolver selects the chronology-maximal conflicting version (Resolve(C_κ)=arg max ν(q)), matching the benchmark's recency-as-truth rule, and the durabl

What carries the argument

The load-bearing mechanism is the Ordered PatchTest (Eq. 1): a deterministic predicate combining a field-completeness check F(q), substring evidence membership Q(e,D_s), and the ordered token-subsequence relation T(v)⪯T(e). It is what converts 'source support' from a fuzzy overlap score into an enforceable commit condition, and it is the only gate in the audit with zero false accepts and zero false rejects. Two companion structures carry the other contracts: the Temporal Resolver (Eq. 2) applies the declared recency-as-truth rule to pick the visible version, and the durable snapshot journal (Eq. 3) persists the application-visible preimage A_0=Π_app(S_0) as the recovery intent, making the ph

Load-bearing premise

The gate's correctness rests on every legitimate update's value being an in-order, verbatim token subsequence of its cited evidence—if real updates require paraphrase, inference, or the evidence is itself false, the gate will either reject the right value or admit the wrong one.

What would settle it

Take a set of real agent update logs where the ground-truth correct value is a paraphrase of the evidence (e.g., evidence 'the flight departs at 14:30' with the correct value 'the flight leaves at 2:30 PM'); if Ordered PatchTest rejects most of these source-supported updates, the gate's claimed support is bounded to verbatim extraction only. Conversely, if even one accepted probe's value is not an ordered subsequence of its cited source, the audit's exactness claim is falsified.

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

If this is right

  • A memory store can reject unsupported writes before they are persisted, so a bad extraction never becomes durable state.
  • When sources conflict, the answer model sees only the chronology-preferred version, improving fact-consolidation F1 over dense retrieval by 17–24 points in the paper's settings.
  • After a persistent fault that survives reopen, the full declared active map is restored, eliminating mixed states that single-key undo leaves behind.
  • Admission, fallback, and rollback stay observable outside the answer model, so governance can be audited and replayed without labels.
  • Because the gate is linear in the inspected token sequences and makes no verifier calls, the contract comes at low computational cost.

Where Pith is reading between the lines

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

  • The exactness of Ordered PatchTest on the audit suggests a strong prior for verbatim evidence in many extraction pipelines; a natural-input test with paraphrased-but-supported values would show where the lexical contract must be relaxed.
  • If the lexical contract is relaxed in future work, the same transaction structure—admission, resolution, recovery—should remain intact, so the paper's contribution may outlive the specific gate.
  • The recency-as-truth resolver will reverse its benefit in domains where the older fact is correct (e.g., superseded policies or revoked permissions); a controlled reverse-chronology variant of the consolidation benchmark would quantify that reversal.
  • Because the gate is model-free and label-free, it can be layered onto any existing memory store; the results suggest a deterministic pre-commit check is worth combining with learned memory operations rather than replacing them.

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

3 major / 4 minor

Summary. The paper proposes MemTxn, an answer-model-external governance layer for writable agent memory. It formalizes three contracts: (i) Ordered PatchTest (Eq. 1) admits an update only if every token of the proposed value appears as an ordered subsequence of a cited evidence span; (ii) a deterministic conflict trigger plus Temporal Resolver (Eq. 2) selects the chronologically newest version under a declared recency-as-truth contract; and (iii) a durable snapshot journal (Eq. 3) restores the complete active-map preimage after persistent multi-key faults. Experiments include an item-disjoint source-support audit with 60 supported and 179 hard-negative probes, 272 recovery runs on LongMemEval-S and LoCoMo states, and MemoryAgentBench FactConsolidation, where MemTxn reports the highest F1 across model configurations and large gains over Dense (17.06–24.07 points in five settings). The paper is transparent about its declared contracts and explicitly disclaims semantic truth, concurrency robustness, and physical-loss recovery.

Significance. If accepted under its declared contracts, MemTxn is a clean and useful separation of write admission, conflict visibility, and recovery in agent memory. The recovery experiments are independently grounded: faults are injected, the database is reopened, and success requires full active-map restoration plus invariant checks, so the 272/272 result is a genuine demonstration of write-set-independent recovery. The FactConsolidation gains are consistent across models and survive a matched-pool control with positive confidence intervals. The paper also ships a deterministic, label-blind, no-verifier-call gate with explicit complexity bounds. The main weakness is that the source-support claim rests on a lexical containment predicate whose natural-language adequacy is not empirically tested: the audit's negatives are template-generated from the same contract, so 100% accuracy is close to a logical consequence. The FactConsolidation result is likewise partly aligned by construction because the benchmark's gold label is recency. These caveats are acknowledged in the Discussion, but the abstract and introduction state the stronger claim that MemTxn 'verifies whether an update is supported by i

major comments (3)
  1. [Source-supported update admission, Eq. (1); Table 1] Eq. (1) equates source support with ordered token-subsequence containment: T(v) ⪯ T(e) ∧ Q(e, D_s). This is both over-inclusive (a false or context-free statement in the cited span passes if its tokens appear in order) and under-inclusive (paraphrase, synonym substitution, cross-span inference, and arithmetic updates are rejected). The Table 1 audit cannot break this symmetry because all 179 negatives are template-generated (negation insertion, token reordering, out-of-source substitution) and the 60 positives are extractor values that are verbatim subsequences of the cited evidence. The 100% accuracy is therefore largely a logical consequence of the predicate, not an empirical finding about natural updates. The Discussion waives 'coverage under naturally occurring update triggers,' which is precisely the gap that determines whether the gate is useful beyond the controlled setting. Pleas
  2. [Temporal Resolution on FactConsolidation, Eq. (2); Tables 2–3] The FactConsolidation benchmark defines the latest conflicting assertion as correct, and Eq. (2) implements exactly that recency rule. Thus the large F1 gains over Dense demonstrate that applying a chronology contract improves this benchmark, but they do not independently validate the Temporal Resolver as a general conflict-resolution method (e.g., when recency is not truth, or when chronology metadata is noisy). The paper states that the benchmark matches the contract, so this is not an internal inconsistency; however, the wording 'achieves the highest average F1' and the abstract's unqualified 'selects the visible version when facts conflict' should be qualified as 'under the benchmark's recency-as-truth contract.'
  3. [Abstract vs. Experiments (configuration count)] The abstract states 'all twelve answer-model configurations,' but Table 2 reports five representative model settings, Table 3 reports six matched-control models, and Figure 6 reports 48 default-protocol cells across six models. The count of twelve and the averaging procedure are not explained anywhere in the experiments. This is a presentation mismatch that should be corrected or clarified.
minor comments (4)
  1. [Figure 3 caption] The caption notes that 'exact checkpoint revisions were not independently verified.' If this is a limitation of the end-to-end workflow measurements, it should be stated in the limitations section and taken into account when interpreting the 1.0/0.992 workflow accuracies.
  2. [Table 4 / Complete-State Recovery] The paper reports zero failures over 272 runs but does not give a per-fault-class breakdown (how many partial-commit, event-divergence, etc., runs make up the 272). A per-class table would make the recovery claim easier to assess.
  3. [Table 1 baselines] The fixed and calibrated coverage baselines are described only by two thresholds (θ=0.60 and θ=0.95). Please specify the calibration procedure and the development set used to choose these thresholds, since the paper stresses that calibration precedes the item-disjoint audit.
  4. [Contract implications] The sentence 'frozen regression tasks remain evaluator-side' is unclear; it would help to explain how this affects reproducibility of the stated guarantees.

Circularity Check

2 steps flagged

FactConsolidation F1 and the 100% source-support audit reduce by construction to the paper's own rules (Eq. 2 recency, Eq. 1 ordered subsequence); recovery and ablations remain independent.

specific steps
  1. self definitional [Experiments — Temporal Resolution on FactConsolidation; Eq. (2)]
    "MemoryAgentBench FactConsolidation (Hu, Wang, and McAuley 2026) defines the latest conflicting assertion as correct, matching the resolver’s chronology contract."

    The benchmark's correctness label is 'latest assertion is correct'; Eq. (2) selects arg max ν(q), i.e., the latest version by declared chronology. MemTxn's F1 on this benchmark therefore measures how well the system implements the benchmark's own ground-truth rule. The high F1 is partly definitional: the resolver was designed to match the label rule, so the experiment validates the implementation, not the choice of recency-as-truth against an independent ground truth. The paper's own scope note ('does not establish semantic truth; it only applies the chronology rule defined by the benchmark or application') confirms the overlap.

  2. self definitional [Experiments — Source-Support Admission; Eq. (1)]
    "Support_ord(q) =F(q)∧Q(e,D s)∧[T(v)⪯T(e)].(1) ... They comprise 60 supported originals and 179 hard negatives generated by negation insertion, token reordering, and out-of-source substitution."

    The 60 positives are extractor values satisfying the ordered-subsequence condition of Eq. (1), and the 179 negatives are template edits (negation insertion, token reordering, out-of-source substitution) that are constructed to break T(v)⪯T(e). Hence the reported 100% acceptance/rejection is a direct application of the predicate to probes generated from the predicate; it is a self-consistency check, not a precision/recall measurement on naturally occurring updates. The paper explicitly scopes this to the declared lexical contract, which mitigates the problem but means the headline 'accepts all... rejects all...' restates the rule rather than independently validating it.

full rationale

The recovery experiments are independently grounded: faults are injected into benchmark-derived states and success requires restoring the full pre-fault active map without the physical write set, a criterion not built into the snapshot mechanism's definition. The ablations and stress tests are also externally varying parameters. The main circularity is concentrated in two headline evaluations. (1) FactConsolidation: the benchmark defines 'latest assertion is correct', and Eq. (2) selects the latest by ν; the F1 improvement is thus a consistency test of the resolver against the benchmark's own label rule. (2) Source-support audit: positives and negatives are constructed around Eq. (1)'s ordered-subsequence condition, so the perfect scores are by construction. No load-bearing self-citation was found; the references to RealMem and other benchmarks are contextual, and no uniqueness theorem is imported from the authors' prior work. The paper is unusually transparent about these boundaries (e.g., 'Coverage under naturally occurring update triggers remains outside the controlled evaluation'), so this is partial circularity rather than a hidden fit. Score 6 reflects that the two headline results reduce by construction while the recovery contribution retains independent content.

Axiom & Free-Parameter Ledger

2 free parameters · 5 axioms · 3 invented entities

The central mechanism is a set of declared contracts, not a derived law. The ledger captures the assumptions those contracts encode: lexical support is the admission criterion, recency is the truth rule, and the snapshot must remain intact. The paper is transparent about these assumptions, but they are real assumptions and the headline results depend on them.

free parameters (2)
  • conflict_trigger_jaccard_threshold = 0.60
    In Conflict-conditioned version resolution, keys with normalized subject-relation Jaccard at least .60 or exact key match trigger conflict detection. The threshold is hand-chosen; no sensitivity analysis or derivation is provided.
  • value_equivalence_jaccard_threshold = 0.80
    A conflict requires values that are neither exact nor Jaccard-equivalent at .80. This threshold is also hand-chosen and not justified by data or theory.
axioms (5)
  • ad hoc to paper The cited evidence span e is authoritative and ordered token containment T(v) ⪯ T(e) is necessary and sufficient for source support.
    Eq. (1) defines support entirely by ordered lexical containment. This is a modeling contract, not a fact about language or memory. Updates requiring paraphrase or cross-span inference are rejected; false but lexically supported values are accepted.
  • ad hoc to paper Chronology metadata ν is reliable and recency-as-truth is the correct conflict-resolution contract.
    Eq. (2) selects arg max ν(q). The paper scopes this to benchmarks/applications that define recency as truth, but it is still an assumed contract, not derived. It matches FactConsolidation's labeling rule.
  • domain assumption The active-map snapshot A0 and recovery intent survive faults and remain intact, and an external detector invokes the controller.
    Eq. (3) assumes PersistIntent(A0) and Restore(A0) work, i.e., the trusted base is intact. Physical loss or intent corruption is explicitly outside scope.
  • domain assumption The upstream extractor provides a well-formed proposal q = (id, κ, u, r, v, e, s, ν).
    The system model assumes the extractor supplies normalized keys, values, evidence spans, source IDs, and chronology. Extraction quality is not evaluated; coverage under naturally occurring update triggers is outside the controlled evaluation.
  • domain assumption SQLite WAL with synchronous=FULL provides atomic durable physical writes.
    The implementation relies on SQLite's atomicity; this is standard storage behavior and is not independently verified in the paper.
invented entities (3)
  • Ordered PatchTest gate no independent evidence
    purpose: Deterministic lexical ordered-subsequence check that admits or rejects proposed memory updates before commit.
    A software component; its behavior is defined and evaluated in this paper, with no external falsifiable prediction beyond the paper's own audits.
  • Temporal Resolver no independent evidence
    purpose: Chooses the visible version among conflicting candidates by maximum chronology metadata.
    Software component; effectively recency-based selection, not a new physical entity. No external handle.
  • Durable snapshot journal no independent evidence
    purpose: Persists the application-visible active map preimage and restores it after a fault.
    Software component; a snapshot journal applied to agent memory. No independent evidence outside the fault-injection experiments.

pith-pipeline@v1.3.0-daily-deepseek · 13901 in / 10916 out tokens · 98879 ms · 2026-08-01T00:31:55.418658+00:00 · methodology

0 comments
read the original abstract

Persistent memory lets long-running large language model agents reuse information across sessions and tasks. Yet errors in writable memory can persist and corrupt future behavior. Existing systems improve storage and retrieval, but they do not provide a transaction boundary for reliable updates and recovery. We therefore propose MemTxn, a governance layer outside the answer model. MemTxn verifies whether an update is supported by its source. It also selects the visible version when facts conflict and restores the application-visible state after a fault. The system uses Ordered PatchTest to validate writes, a Temporal Resolver to select versions, and a durable snapshot journal to recover state. On an item-disjoint audit, MemTxn accepts all 60 supported originals and rejects all 179 hard negatives. Under persistent multi-key faults on LongMemEval-S and LoCoMo states, it restores the complete declared active map without knowing the actual physical write set. On MemoryAgentBench FactConsolidation, MemTxn achieves the highest average F1 across all twelve answer-model configurations. It outperforms Dense by 17.06--24.07 points in five representative settings.

Figures

Figures reproduced from arXiv: 2607.27834 by Fanshuai Meng, Hanshuai Cui, Qianli Ma, Weijia Jia, Zhiqing Tang, Zhi Yao.

Figure 1
Figure 1. Figure 1: MemTxn rejects unsupported edits before commit [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: MemTxn workflow on the return-policy example. The write lane governs activation, the answer lane resolves conflicts, [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Probe-class source-support decisions. Correct [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Matched-budget FactConsolidation trace with one [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: MemTxn–Dense F1 gains across models and hop/context-length settings. zero failures over 272 runs, whereas the single-key baseline fails all 136 partial-commit and event-divergence runs. Thus MemTxn matches the privileged oracle’s coverage without receiving the physical write set. No design falsely rolls back [PITH_FULL_IMAGE:figures/full_fig_p006_6.png] view at source ↗
Figure 8
Figure 8. Figure 8: Robustness to gate reliability and sensitivity axes. [PITH_FULL_IMAGE:figures/full_fig_p007_8.png] view at source ↗
Figure 7
Figure 7. Figure 7: Component ablations and update-scale accuracy. [PITH_FULL_IMAGE:figures/full_fig_p007_7.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

40 extracted references · 4 linked inside Pith

  1. [1]

    Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology , pages =

    Generative Agents: Interactive Simulacra of Human Behavior , author =. Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology , pages =. 2023 , doi =

    2023

  2. [2]

    and Stoica, Ion and Gonzalez, Joseph E

    Packer, Charles and Wooders, Sarah and Lin, Kevin and Fang, Vivian and Patil, Shishir G. and Stoica, Ion and Gonzalez, Joseph E. , year =. 2310.08560 , archivePrefix =

    Pith/arXiv arXiv

  3. [3]

    Advances in Neural Information Processing Systems , volume =

    Reflexion: Language Agents with Verbal Reinforcement Learning , author =. Advances in Neural Information Processing Systems , volume =. 2023 , url =

    2023

  4. [4]

    Transactions on Machine Learning Research , year =

    Voyager: An Open-Ended Embodied Agent with Large Language Models , author =. Transactions on Machine Learning Research , year =

  5. [5]

    Mem0: Building Production-Ready

    Chhikara, Prateek and Khant, Dev and Aryan, Saket and Singh, Taranjeet and Yadav, Deshraj , booktitle =. Mem0: Building Production-Ready. 2025 , doi =

    2025

  6. [6]

    2025 , url =

    Xu, Wujiang and Liang, Zujie and Mei, Kai and Gao, Hang and Tan, Juntao and Zhang, Yongfeng , booktitle =. 2025 , url =

    2025

  7. [7]

    Kang, Jiazheng and Ji, Mingming and Zhao, Zhe and Bai, Ting , booktitle =. Memory. 2025 , doi =

    2025

  8. [8]

    2507.03724 , archivePrefix =

    Li, Zhiyu and Xi, Chenyang and Li, Chunyu and Chen, Ding and Chen, Boyu and Song, Shichao and Niu, Simin and Wang, Hanyu and Yang, Jiawei and Tang, Chen and Yu, Qingchen and Zhao, Jihao and Wang, Yezhaohui and Liu, Peng and Lin, Zehao and Wang, Pengyuan and Huo, Jiahao and Chen, Tianyi and Chen, Kai and Li, Kehang and Tao, Zhen and Lai, Huayi and Wu, Hao ...

    Pith/arXiv arXiv

  9. [9]

    2026 , url =

    Fang, Jizhan and Deng, Xinle and Xu, Haoming and Jiang, Ziyan and Tang, Yuqi and Xu, Ziwen and Deng, Shumin and Yao, Yunzhi and Wang, Mengru and Qiao, Shuofei and Chen, Huajun and Zhang, Ningyu , booktitle =. 2026 , url =

    2026

  10. [10]

    2025 , howpublished =

    2025

  11. [11]

    2025 , eprint =

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

    2025

  12. [12]

    From Human Memory to

    Wu, Yaxiong and Liang, Sheng and Zhang, Chen and Wang, Yichao and Zhang, Yongyue and Guo, Huifeng and Tang, Ruiming and Liu, Yong , year =. From Human Memory to. doi:10.48550/arXiv.2504.15965 , url =. 2504.15965 , archivePrefix =

    Pith/arXiv arXiv doi:10.48550/arxiv.2504.15965

  13. [13]

    From Storage to Experience: A Survey on the Evolution of

    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 , booktitle =. From Storage to Experience: A Survey on the Evolution of. 2026 , doi =

    2026

  14. [14]

    How Memory Management Impacts

    Xiong, Zidi and Lin, Yuping and Xie, Wenya and He, Pengfei and Liu, Zirui and Tang, Jiliang and Lakkaraju, Himabindu and Xiang, Zhen , booktitle =. How Memory Management Impacts. 2026 , doi =

    2026

  15. [15]

    Agent-Memory Protocol: A Privacy-Focused Protocol for

    Wu, Junde and Hu, Minhao and Zhu, Jiayuan and Wang, Jiaye and Jin, Yueming , booktitle =. Agent-Memory Protocol: A Privacy-Focused Protocol for. 2026 , url =

    2026

  16. [16]

    Retrieval-Augmented Generation for Knowledge-Intensive

    Lewis, Patrick and Perez, Ethan and Piktus, Aleksandra and Petroni, Fabio and Karpukhin, Vladimir and Goyal, Naman and K. Retrieval-Augmented Generation for Knowledge-Intensive. Advances in Neural Information Processing Systems , volume =. 2020 , url =

    2020

  17. [17]

    Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP) , pages =

    Dense Passage Retrieval for Open-Domain Question Answering , author =. Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP) , pages =. 2020 , doi =

    2020

  18. [18]

    2024 , doi =

    Chen, Jianlyu and Xiao, Shitao and Zhang, Peitian and Luo, Kun and Lian, Defu and Liu, Zheng , booktitle =. 2024 , doi =

    2024

  19. [19]

    2025 , url =

    Wu, Di and Wang, Hongwei and Yu, Wenhao and Zhang, Yuwei and Chang, Kai-Wei and Yu, Dong , booktitle =. 2025 , url =

    2025

  20. [20]

    Evaluating Very Long-Term Conversational Memory of

    Maharana, Adyasha and Lee, Dong-Ho and Tulyakov, Sergey and Bansal, Mohit and Barbieri, Francesco and Fang, Yuwei , booktitle =. Evaluating Very Long-Term Conversational Memory of. 2024 , doi =

    2024

  21. [21]

    Evaluating Memory in

    Hu, Yuanzhe and Wang, Yu and McAuley, Julian , booktitle =. Evaluating Memory in. 2026 , url =

    2026

  22. [22]

    Ross , booktitle =

    Tavakoli, Mohammad and Salemi, Alireza and Ye, Carrie and Abdalla, Mohamed and Zamani, Hamed and Mitchell, J. Ross , booktitle =. Beyond a Million Tokens: Benchmarking and Enhancing Long-Term Memory in. 2026 , url =

    2026

  23. [23]

    2025 , doi =

    Tan, Haoran and Zhang, Zeyu and Ma, Chen and Chen, Xu and Dai, Quanyu and Dong, Zhenhua , booktitle =. 2025 , doi =

    2025

  24. [24]

    2026 , doi =

    Shen, Yiting and Li, Kun and Zhou, Wei and Hu, Songlin , booktitle =. 2026 , doi =

    2026

  25. [25]

    2026 , doi =

    Bian, Haonan and Yao, Zhiyuan and Hu, Sen and Xu, Zishan and Zhang, Shaolei and Guo, Yifu and Yang, Ziliang and Han, Xueran and Wang, Huacan and Chen, Ronghao , booktitle =. 2026 , doi =

    2026

  26. [26]

    Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing , pages =

    Editing Factual Knowledge in Language Models , author =. Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing , pages =. 2021 , doi =

    2021

  27. [27]

    International Conference on Learning Representations , year =

    Fast Model Editing at Scale , author =. International Conference on Learning Representations , year =

  28. [28]

    Locating and Editing Factual Associations in

    Meng, Kevin and Bau, David and Andonian, Alex and Belinkov, Yonatan , booktitle =. Locating and Editing Factual Associations in. 2022 , url =

    2022

  29. [29]

    International Conference on Learning Representations , year =

    Mass-Editing Memory in a Transformer , author =. International Conference on Learning Representations , year =

  30. [30]

    2024 , doi =

    Wang, Peng and Li, Zexi and Zhang, Ningyu and Xu, Ziwen and Yao, Yunzhi and Jiang, Yong and Xie, Pengjun and Huang, Fei and Chen, Huajun , booktitle =. 2024 , doi =

    2024

  31. [31]

    2025 , url =

    Fang, Junfeng and Jiang, Houcheng and Wang, Kun and Ma, Yunshan and Shi, Jie and Wang, Xiang and He, Xiangnan and Chua, Tat-Seng , booktitle =. 2025 , url =

    2025

  32. [32]

    Yan, Sikuan and Yang, Xiufeng and Huang, Zuchao and Nie, Ercong and Ding, Zifeng and Li, Zonggen and Ma, Xiaowen and Bi, Jinhe and Kersting, Kristian and Pan, Jeff Z. and Sch. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages =. 2026 , doi =

    2026

  33. [33]

    2024 , eprint =

    A Comprehensive Study of Knowledge Editing for Large Language Models , author =. 2024 , eprint =. doi:10.48550/arXiv.2401.01286 , url =

    Pith/arXiv arXiv doi:10.48550/arxiv.2401.01286 2024

  34. [34]

    Advances in Neural Information Processing Systems , volume =

    Toolformer: Language Models Can Teach Themselves to Use Tools , author =. Advances in Neural Information Processing Systems , volume =. 2023 , url =

    2023

  35. [35]

    and Hashimoto, Tatsunori , booktitle =

    Ruan, Yangjun and Dong, Honghua and Wang, Andrew and Pitis, Silviu and Zhou, Yongchao and Ba, Jimmy and Dubois, Yann and Maddison, Chris J. and Hashimoto, Tatsunori , booktitle =. Identifying the Risks of. 2024 , url =

    2024

  36. [36]

    Advances in Neural Information Processing Systems , volume =

    Debenedetti, Edoardo and Zhang, Jie and Balunovi. Advances in Neural Information Processing Systems , volume =. 2024 , doi =

    2024

  37. [37]

    2025 , url =

    Yao, Shunyu and Shinn, Noah and Razavi, Pedram and Narasimhan, Karthik , booktitle =. 2025 , url =

    2025

  38. [38]

    Agent Security Bench (

    Zhang, Hanrong and Huang, Jingyuan and Mei, Kai and Yao, Yifei and Wang, Zhenting and Zhan, Chenlu and Wang, Hongwei and Zhang, Yongfeng , booktitle =. Agent Security Bench (. 2025 , url =

    2025

  39. [39]

    and Haderle, Don and Lindsay, Bruce and Pirahesh, Hamid and Schwarz, Peter , journal =

    Mohan, C. and Haderle, Don and Lindsay, Bruce and Pirahesh, Hamid and Schwarz, Peter , journal =. 1992 , doi =

    1992

  40. [40]

    2026 , howpublished =

    Write-Ahead Logging , author =. 2026 , howpublished =

    2026