Pith.

Artificial Intelligence · arXiv:2610.21170 v1 · submitted 15 Oct 2026 · 74 references

MemScope: Evolving Agent Memory Architectures through Localized Code Repair

Miao Su, Yucan Guo, Zixuan Li, Long Bai, Zhongni Hou, Xiaolong Jin, Jiafeng Guo

The paper says

“Agent memory architectures define how LLM agents store and reuse experience across tasks, but they are typically fixed after deployment, limiting their adaptability across task types.”

more of the abstract

Recent self-evolving approaches instead adapt these memory architectures using execution feedback, typically represented as evaluation scores or compressed summaries. However, such coarse-grained feedback is insufficient for precise failure localization in current agent systems, where long-horizon task execution involves multiple interacting components. Without a precise failure location, existing approaches lack a clear code-level target for modification, leaving the update scope weakly constrained and often leading to large-scale rewrites of an otherwise capable memory architecture. We propose MemScope, a framework for evidence-grounded, localized evolution of an existing memory architecture. Its tool-using Patch Agent jointly inspects raw execution traces, runtime memory state, and current architecture code to identify repair targets and apply bounded, in-place edits under provider-identity and memory-compatibility constraints. On GAIA, the trace-and-memory inspection variant achieves 86.67% accuracy, compared with 70.00% for trace-only and 63.33% for summary-only feedback. Under xBench tasks, the evolved memory architecture improves over its original version from 49 to 55 correct answers with DeepSeek V3.2 and from 50 to 53 with Kimi K2.7 Code. On disjoint ALFWorld unseen tasks, success rises from 50.75% to 71.64%. Across archived architecture-snapshot transitions, the median code-change ratio is 4.4%. These results support fine-grained execution evidence as a basis for localized, in-place memory-architecture evolution.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 16 Oct 2026 · what was read was not recorded · not checked by a person

It can be wrong. Challenge this review or reply to it.

Holds up, with conditions. Solid paper with a clear idea—localized memory-architecture repair via execution traces—but the stress test concern is real: they never validate that the Patch Agent actually localizes correctly, so the claimed mechanism is inferred from task accuracy rather than measured directly. This line is wrong

3 major objections · 5 minor

How often objections from reviews like this one are right: not yet measured

Are you an author? What you can do.

Not checked: figures, code, what the 74 references say.

The paper says, at §3.3, Table 5

The review did not quote this passage.

Pith says: Objection 1 of 3 · major

The central claim is that fine-grained execution evidence enables localized, evidence-grounded repair. However, the evaluation does not directly measure whether the Patch Agent correctly identifies the root-cause code region. The GAIA diagnostic ablation (Table 5) varies evidence sources and reports final task accuracy, which conflates diagnosis quality with repair efficacy.

more of objection 1

Without a direct metric of localization accuracy (e.g., fraction of edits that match a ground-truth bug location, or human evaluation of Patch Agent diagnoses), the claimed mechanism remains unvalidated. The paper's limitation statement (§B.5) discusses cost but not this issue. The authors should either provide such a direct evaluation or explicitly acknowledge this as a limitation and discuss alternative explanations for the observed performance gains (e.g., serendipitous changes).

The paper says, at §3.2, repair trigger

The review did not quote this passage.

Pith says: Objection 2 of 3 · major

The repair gate uses two free parameters, δ=0.05 and f_min=1, without sensitivity analysis. The results could be sensitive to these thresholds; for instance, a higher failure threshold might yield fewer patches and different outcomes. The authors should report the effect of varying these parameters on the number of patches triggered, final accuracy, or code-change ratio, or else justify the choices more rigorously.

The paper says, at §4.1, embedding model

The review did not quote this passage.

Pith says: Objection 3 of 3 · major

The embedding model all-MiniLM-L6-v2 is used for semantic retrieval throughout. Since several repairs modify retrieval thresholds and filtering logic, the choice of embedding model could affect both diagnosis quality and final performance. No ablation or discussion of embedding model sensitivity is provided. The authors should either test alternative embeddings or discuss this as a potential confound.

5 minor objections

The paper says, at Abstract, §4.4, §5

Pith says: minor objection 1 of 5

The phrase 'trace-and-memory inspection achieves 86.67% accuracy' (used in multiple places) is ambiguous: it refers to task accuracy on a GAIA diagnostic ablation, not diagnostic accuracy. Please explicitly state the metric (task accuracy) each time to avoid confusion.

The paper says, at Table 3

Pith says: minor objection 2 of 5

The paired outcome table is informative, but it would be strengthened by a statistical test (e.g., McNemar's test) to quantify the significance of the improvement and the relative rates of recovered failures versus regressions.

show 3 more

The paper says, at §B.5

Pith says: minor objection 3 of 5

The evolution cost analysis is qualitative. If possible, include approximate token counts or API costs for the diagnosis, code-modification, and validation stages to support the claim that localized repair is more efficient than complete generation.

The paper says, at Figure 3

Pith says: minor objection 4 of 5

The guidance-coverage and conditional-accuracy plot would benefit from error bars or confidence intervals, especially since the online trajectory differs between original and evolved providers (memory state also varies).

The paper says, at §3.3, Table 7

Pith says: minor objection 5 of 5

It would be helpful to report the frequency of no_patch decisions across all runs, to give a sense of how often the Patch Agent judges the evidence insufficient for a change.

What would settle it

Sample the GAIA diagnostic ablation episodes (30 tasks) and have a human evaluator (or a second, independently prompted LLM) annotate, for each patch attempt, whether the Patch Agent's identified repair target (the function or code block edited) matches the true root cause of the failure as determined from the full execution trace and the provider code. Compute the proportion of correctly localized patches. If this proportion is at or below 50%, the claimed localization mechanism is not supported and the performance gains may be due to other factors.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 16 Oct 2026

  1. Reader holds up, with conditions confidence high
  2. Skeptic partly agrees 41 s · 20,556 tokens read, 4,166 written
  3. Editor wrote the verdict above 11 s · 21,473 tokens read, 1,024 written
  4. Referee major revision: 3 major objections, 5 minor 21 s · 20,825 tokens read, 2,419 written

Not run on this paper: the simulated authors' defence.

Pith did not check this review's objections against the paper.

The review in full

The review in brief, machine-written

The paper's core idea is clean and motivated. Prior work (MemEvolve, ALMA) feeds coarse feedback to an LLM that rewrites the whole memory provider. MemScope instead gives the Patch Agent raw traces, memory state, and provider code, then asks for a targeted, in-place edit. This makes failure localization explicit, which is genuinely sensible. The empirical picture is reasonable: GAIA improves substantially (+15–23 pp), ALFWorld shows a clear gain (50.75% → 71.64%), and the ablation (Table 5) shows that trace+memory > trace-only > summary-only. The xBench gains are modest (6 and 3 out of 70) but the matched-start comparison is fair. They also document the edit locality (median 4.4% change) which supports the claim of bounded modification.

The soft spots are real but not fatal. The big one is the stress-test concern: the paper never measures whether the Patch Agent identifies the correct code region. Task accuracy conflates diagnosis quality with repair efficacy. The ablation varies evidence sources and only measures final accuracy, not diagnostic correctness. The case studies (Table 6) are anecdotal. The paper's central claim—that fine-grained execution evidence enables localized, targeted repair—rests on an unvalidated step. They could add a direct localization metric (e.g., how often the edited region matches a human-judged root cause). The other limitations are secondary: one starting architecture (DynamicCheatsheet), one agent harness (Flash-Searcher for web tasks), no error bars, no code release. The paper is thorough in other respects (ablation, failure retention analysis, patch history tracking), so these are more about generality than internal validity.

The tone should be proportionate: this is a solid first paper on a useful idea, not a breakthrough. The central claim is plausible and supported by the evidence they provide, but the evidence is indirect. A serious referee would ask for direct validation of the diagnostic step.

Recommendation: I would accept for peer review, but I'd push for a direct localization accuracy metric. The paper is well-written, the method is clearly described, and the results are positive. The lack of code and the single-architecture issue are standard limitations that don't invalidate the claim. The stress-test concern is the one that weakens the core mechanism claim, but it doesn't undermine the overall finding that trace+memory feedback improves evolution outcomes.

The strongest objection

The central claim that the Patch Agent 'jointly inspects raw execution traces, runtime memory state, and current provider code to identify repair targets' (§3.3) is only indirectly supported. The ablation (Table 5) varies evidence sources and measures final task accuracy, but never measures whether the Patch Agent correctly identifies the root-cause code region. The case studies (Table 6, xBench examples) are illustrative but not systematically quantified. The median code-change ratio of 4.4% (Table 8) shows edits are small, but does not indicate whether the edited region is the correct one. If the Patch Agent frequently localizes incorrectly, the observed performance gains could arise from serendipitous changes rather than targeted repair, weakening the paper's claimed contribution—that fine-grained execution evidence enables localized evolution. The paper's own limitation statement (§B.5) discusses cost but not validation of diagnostic accuracy. Without a direct localization metric, the argument that 'trace-and-memory inspection achieves 86.67% accuracy' (Figure 1, sentence [b6]) is a claim about task performance, not about diagnostic correctness.

The reader's weakest_assumption—that the Patch Agent can reliably distinguish bug types—is closely related but not identical. My concern is narrower: the paper never measures whether the Patch Agent's localization is correct, only whether the final provider performs better. This is a load-bearing gap because the paper's novelty hinges on making 'failure localization an explicit step' (sentence [b40]), yet no experiment validates that the localization step actually picks the right code region. I still find the empirical results strong enough to support the verdict CONDITIONAL, as the overarching claim (that the method improves performance) is well-supported. However, the mechanistic claim about localization is weaker. I recommend retaining the CONDITIONAL verdict with a recommendation to add a direct localization accuracy analysis in the revision, either via human evaluation or a synthetic dataset with known bugs. The reader's rationale already notes 'no code or data are released' and 'no error bars', which are related but distinct; my concern provides a concrete, targeted check that would strengthen the paper's central argument.

The referee's report

If the central claim holds, MemScope demonstrates a principled method for adapting agent memory architectures after deployment, addressing a practical limitation of fixed architectures. The localized repair approach could reduce the risks of large-scale rewrites and make self-evolution more efficient. The experimental setup includes multiple benchmarks, backbone models, and ablation conditions, providing broad coverage. Strengths include the use of a diverse starting architecture (DynamicCheatsheet), matched-start comparisons, and an established evaluation protocol following prior work (MemEvolve, ALMA). However, the paper's significance is tempered by a missing direct validation of the diagnostic mechanism.

Pith's reading of the paper (a machine's summary, not the authors')

The paper argues that to improve an LLM agent's memory architecture, one should not rewrite it wholesale based on outcome scores, but instead inspect exactly what went wrong during task execution. By giving a repair system access to raw execution traces, runtime memory records, and the provider's code, MemScope identifies the smallest code region responsible for a failure and applies a bounded, in-place edit. This approach preserves working functionality while fixing specific defects. Across benchmarks, it outperforms prior methods that rely on evaluation scores or compressed summaries.

What carries the argument. MemScope's central mechanism is the Patch Agent, an LLM equipped with read-only tools for trace inspection, memory-state querying, and provider-code reading. Given a set of failed and successful task executions, the Patch Agent uses these tools to identify the exact code region responsible for a recurring failure pattern, proposes an old-to-new replacement under identity and compatibility constraints, and validates the edit locally. The repair is bounded to avoid broad rewrites: the median transition touches 4.4% of the provider's lines and a single function.

Core claim. The central discovery is that fine-grained execution evidence—raw traces of task steps, the runtime memory state showing what was stored and retrieved, and the current provider code—can be jointly used to localize a defect to a single function or threshold parameter in an agent's memory architecture. This turns evolution from a complete-architecture generation problem into a targeted code-repair problem. The paper shows that this localization is accurate enough to produce consistent performance gains with a median code-change ratio of 4.4% per transition, and that trace-and-memory inspection achieves 86.67% diagnostic accuracy compared to 63.33% for summary-only feedback.

The premise it rests on. The load-bearing premise is that the Patch Agent (an LLM) can reliably distinguish memory-storage bugs, retrieval bugs, and downstream agent bugs using only the raw traces, runtime memory state, and provider code it inspects, without needing to run the full benchmark.

If the paper is right (Pith's inferences, not the authors')

If MemScope is correct, future self-evolving agents can adopt a similar local-repair loop instead of generating complete candidate architectures, reducing the risk of regressions and the cost of evaluation.

The approach implies that memory-performance diagnostics can be largely automated: an LLM with targeted inspection tools can identify the root cause of a retrieval failure, storage error, or formatting defect without executing the full benchmark.

The method suggests that provider-identity and memory-compatibility constraints are sufficient to keep in-place edits safe, enabling continued runtime operation without resetting the memory bank.

The ablation results imply that summary-only feedback is insufficient for localization; trace and memory inspection are complementary evidence sources, with memory-state inspection providing additional diagnostic power beyond raw traces.

The Patch Agent's reliance on an LLM to interpret traces and propose code edits means its diagnostic accuracy is bounded by the LLM's reasoning capabilities; future work could test whether smaller or specialized models achieve similar localization accuracy at lower cost.

MemScope currently starts from a single architecture (DynamicCheatsheet); if the method generalizes to other starting architectures, the evidence-grounded repair principle might unify a broader class of memory-system improvements.

The paper does not explore whether localized repairs accumulate over many episodes into emergent architectural improvements that diverge qualitatively from the starting design; tracking the long-term trajectory of edits could reveal whether local repairs converge or drift.

Check this review

Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.

How often reviews like this one are right

Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.

How often objections from reviews like this one are right: not yet measured

Every value above, in the review's public record (JSON)

Reviews by people

This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.

Request a human review

How Pith checks papers · The full audit of this paper

The paper

Abstract

Agent memory architectures define how LLM agents store and reuse experience across tasks, but they are typically fixed after deployment, limiting their adaptability across task types. Recent self-evolving approaches instead adapt these memory architectures using execution feedback, typically represented as evaluation scores or compressed summaries. However, such coarse-grained feedback is insufficient for precise failure localization in current agent systems, where long-horizon task execution involves multiple interacting components. Without a precise failure location, existing approaches lack a clear code-level target for modification, leaving the update scope weakly constrained and often leading to large-scale rewrites of an otherwise capable memory architecture. We propose MemScope, a framework for evidence-grounded, localized evolution of an existing memory architecture. Its tool-using Patch Agent jointly inspects raw execution traces, runtime memory state, and current architecture code to identify repair targets and apply bounded, in-place edits under provider-identity and memory-compatibility constraints. On GAIA, the trace-and-memory inspection variant achieves 86.67% accuracy, compared with 70.00% for trace-only and 63.33% for summary-only feedback. Under xBench tasks, the evolved memory architecture improves over its original version from 49 to 55 correct answers with DeepSeek V3.2 and from 50 to 53 with Kimi K2.7 Code. On disjoint ALFWorld unseen tasks, success rises from 50.75% to 71.64%. Across archived architecture-snapshot transitions, the median code-change ratio is 4.4%. These results support fine-grained execution evidence as a basis for localized, in-place memory-architecture evolution.

The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.

abstract compiled · 2026-10-16

The authors' main claim, as the classifier picked it out 1.00

“These results support fine-grained execution evidence as a basis for localized, in-place memory-architecture evolution.”

How the main claim is stated: asserted 0.64

What the abstract itself states, as probabilities: claims first 0.06 · states limitation 0.13 · human participants 0.04 · controlled comparison 0.94 · randomized assignment 0.02 · states quantitative result 0.99 · compares with prior results 0.97 · mentions code or data availability 0.04

The abstract, sentence by sentence

This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.

SentenceRoleThe authors' own workStates the main conclusionStrength
Agent memory architectures define how LLM agents store and reuse experience across tasks, but they are typically fixed after deployment, limiting their adaptability across task types. background 0.98 0.04 0.02 asserted 1.00
Recent self-evolving approaches instead adapt these memory architectures using execution feedback, typically represented as evaluation scores or compressed summaries. background 1.00 0.04 0.03 asserted 1.00
However, such coarse-grained feedback is insufficient for precise failure localization in current agent systems, where long-horizon task execution involves multiple interacting components. gap 0.84 0.10 0.03 asserted 1.00
Without a precise failure location, existing approaches lack a clear code-level target for modification, leaving the update scope weakly constrained and often leading to large-scale rewrites of an otherwise capable memory architecture. gap 0.89 0.07 0.04 asserted 0.97
We propose MemScope, a framework for evidence-grounded, localized evolution of an existing memory architecture. contribution 1.00 0.82 0.10 asserted 0.99
Its tool-using Patch Agent jointly inspects raw execution traces, runtime memory state, and current architecture code to identify repair targets and apply bounded, in-place edits under provider-identity and memory-compatibility constraints. method 0.98 0.83 0.07 asserted 1.00
On GAIA, the trace-and-memory inspection variant achieves 86.67% accuracy, compared with 70.00% for trace-only and 63.33% for summary-only feedback. result 1.00 0.96 0.10 asserted 1.00
Under xBench tasks, the evolved memory architecture improves over its original version from 49 to 55 correct answers with DeepSeek V3.2 and from 50 to 53 with Kimi K2.7 Code. result 1.00 0.95 0.06 asserted 1.00
On disjoint ALFWorld unseen tasks, success rises from 50.75% to 71.64%. result 1.00 0.95 0.05 asserted 1.00
Across archived architecture-snapshot transitions, the median code-change ratio is 4.4%. result 1.00 0.95 0.05 asserted 1.00
These results support fine-grained execution evidence as a basis for localized, in-place memory-architecture evolution. interpretation 1.00 0.73 0.89 asserted 0.64

What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.98

What was examined, and what was not

Sources read

  • arXiv listing abstract v1

Not examined

  • full text
  • figures and tables
  • references and citing sentences
  • relations to other papers

Levels run

  • P abstract

Not run

  • S sections
  • T body sentences
  • C citations
  • R cross-paper relations

How the labels were made

Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-10-16: 1 request, 8,346 input tokens.

Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON

Figures (3)
Figure 1 of the paper
Figure 1. Feedback and update granularity in memory-architecture evolution. Prior architecture-level approaches use outcome-level feedback to propose complete executable archi￾tectures. MemScope makes the repair target explicit by jointly inspecting execution traces, runtime memory state, and provider code, then applies a localized, in-place edit to the existing provider. In the illustrated update, only the highlighted code region is modified. is retrieved, and how retrieved information is presented (Zhang et al., 2026a; Xiong et al., 2026a). This shifts the evolution target from memory…
Figure 2 of the paper
Figure 2. Overview of MemScope. (a) Evolution Target. Following the Agent Skills format, each
Figure 3 of the paper
Figure 3. GAIA guidance during online memory updates. Original versus evolved provider coverage:

Figures are the authors', reproduced from the paper.

References (74)
  1. [1] Bengio, Yoshua and LeCun, Yann , booktitle =. Scaling Learning Algorithms Towards on Pith
  2. [2] Hinton, Geoffrey E. and Osindero, Simon and Teh, Yee Whye , journal =. A Fast Learning Algorithm for Deep Belief Nets , volume = on Pith
  3. [3] Deep learning , author=. 2016 , publisher= on Pith
  4. [4] Speculations Concerning the First Ultraintelligent Machine , author=. Advances in Computers , volume=. 1965 , publisher= on Pith
  5. [5] Zelikman, Eric and Lorch, Eliana and Mackey, Lester and Kalai, Adam Tauman , booktitle=. Self-Taught Optimizer (. 2024 , eprint= on Pith
  6. [6] Zhang, Jenny and Hu, Shengran and Lu, Cong and Lange, Robert and Clune, Jeff , journal=. Darwin. 2025 , doi= on Pith
  7. [7] Harness Engineering for Self-Improvement , author=. Lil'Log , year= on Pith
  8. [8] Meta-Harness: End-to-End Optimization of Model Harnesses , author=. 2026 , eprint= on Pith
  9. [9] Self-Harness: Harnesses That Improve Themselves , author=. 2026 , eprint= on Pith
  10. [10] Agentic Harness Engineering: Observability-Driven Automatic Evolution of Coding-Agent Harnesses , author=. 2026 , eprint= on Pith
  11. [11] Reflexion: Language Agents with Verbal Reinforcement Learning , author=. Advances in Neural Information Processing Systems , volume=. 2023 , doi= on Pith
  12. [12] Contextual Experience Replay for Self-Improvement of Language Agents , author=. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=. 2025 , doi= on Pith
  13. [13] Mirac Suzgun and Mert Yuksekgonul and Federico Bianchi and Dan Jurafsky and James Zou , booktitle=. 2026 , doi= on Pith
  14. [14] Guibin Zhang and Haotian Ren and Chong Zhan and Junhao Wang and He Zhu and Wangchunshu Zhou and Shuicheng Yan , booktitle=. 2026 , url= on Pith
  15. [15] 2023 , howpublished= on Pith
  16. [16] Laws, Joe , year=. Trase Tops on Pith
  17. [17] Introducing Deep Research , author=. 2025 , howpublished= on Pith
  18. [18] 2024 , howpublished= on Pith
  19. [19] Desearch: Web Search. 2024 , howpublished= on Pith
  20. [20] Hu, Mengkang and Zhou, Yuhang and Fan, Wendong and Nie, Yuzhou and Xia, Bowei and Sun, Tao and Ye, Ziyu and Jin, Zhaoxuan and Li, Yingru and Chen, Qiguang and Zhang, Zeyu and Wang, Yifeng and Ye, Qianshuo and Ghanem, Bernard and Luo, Ping and Li, Guohao , year=. 2505.23885 , a... on PithPith's reviewarXiv
show 54 more
  1. [21] TapeAgents: A Holistic Framework for Agent Development and Optimization , author=. 2024 , eprint= on Pith
  2. [22] Tang, Jiabin and Fan, Tianyu and Huang, Chao , booktitle=. AutoAgent: A Fully-Automated and Zero-Code Framework for. 2026 , doi= on Pith
  3. [23] Magentic-One: A Generalist Multi-Agent System for Solving Complex Tasks , author=. 2024 , eprint= on Pith
  4. [24] Cognitive Kernel-Pro: A Framework for Deep Research Agents and Agent Foundation Models Training , author=. 2025 , eprint= on Pith
  5. [25] OAgents: An Empirical Study of Building Effective Agents , author=. Findings of the Association for Computational Linguistics: EMNLP 2025 , pages=. 2025 , doi= on Pith
  6. [26] Tang, Xiangru and Qin, Tianrui and Peng, Tianhao and Zhou, Ziyang and Shao, Daniel and Du, Tingting and Wei, Xinming and Xia, Peng and Wu, Fang and Zhu, He and Zhang, Ge and Liu, Jiaheng and Wang, Xingyao and Hong, Sirui and Wu, Chenglin and Cheng, Hao and Wang, Chi and Zhou, ... on PitharXiv
  7. [27] Alita: Generalist Agent Enabling Scalable Agentic Reasoning with Minimal Predefinition and Maximal Self-Evolution , author=. 2025 , eprint= on Pith
  8. [28] Jiang, Xun and Li, Feng and Zhao, Han and Qiu, Jiahao and Wang, Jiaying and Shao, Jun and Xu, Shihao and Zhang, Shu and Chen, Weiling and Tang, Xavier and Chen, Yize and Wu, Mengyue and Ma, Weizhi and Wang, Mengdi and Chen, Tianqiao , year=. Long Term Memory: The Foundation of... on PithPith's reviewarXiv
  9. [29] MemSkill: Learning and Evolving Memory Skills for Self-Evolving Agents , author=. 2026 , eprint= on Pith
  10. [30] Learning to Continually Learn via Meta-learning Agentic Memory Designs , author=. 2026 , eprint= on Pith
  11. [31] Mialon, Gr. The Twelfth International Conference on Learning Representations , year= on Pith
  12. [32] Wu, Jialong and Yin, Wenbiao and Jiang, Yong and Wang, Zhenglin and Xi, Zekun and Fang, Runnan and Zhang, Linhai and He, Yulan and Zhou, Deyu and Xie, Pengjun and Huang, Fei , booktitle=. WebWalker: Benchmarking. 2025 , publisher=. doi:10.18653/v1/2025.acl-long.508 , url= on Pithdoi:10.18653/v1/2025.acl-long.508
  13. [33] Qin, Tianrui and Chen, Qianben and Wang, Sinuo and Xing, He and Zhu, King and Zhu, He and Shi, Dingfeng and Liu, Xinxin and Zhang, Ge and Liu, Jiaheng and Jiang, Yuchen Eleanor and Gao, Xitong and Zhou, Wangchunshu , journal=. Flash-Searcher: Fast and Effective Web Agents via on Pith
  14. [34] Beyond Dialogue Time: Temporal Semantic Memory for Personalized LLM Agents , author=. Findings of the Association for Computational Linguistics: ACL 2026 , pages=. 2026 , doi= on Pith
  15. [35] MemoryBank: Enhancing Large Language Models with Long-Term Memory , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=. 2024 , doi= on Pith
  16. [36] Mem0: Building Production-Ready AI Agents with Scalable Long-Term Memory , author=. ECAI 2025 , publisher=. 2025 , doi= on Pith
  17. [37] Pre-Storage Reasoning for Episodic Memory: Shifting Inference Burden to Memory for Personalized Dialogue , author=. 2025 , eprint= on Pith
  18. [38] Inside Out: Evolving User-Centric Core Memory Trees for Long-Term Personalized Dialogue Systems , author=. 2026 , eprint= on Pith
  19. [39] Agent Workflow Memory , author=. Proceedings of the 42nd International Conference on Machine Learning , series=. 2025 , url= on Pith
  20. [40] Memp: Exploring Agent Procedural Memory , author=. Findings of the Association for Computational Linguistics: ACL 2026 , pages=. 2026 , doi= on Pith
  21. [41] AriGraph: Learning Knowledge Graph World Models with Episodic Memory for LLM Agents , author=. 2025 , eprint= on Pith
  22. [42] How Memory Management Impacts LLM Agents: An Empirical Study of Experience-Following Behavior , author=. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=. 2026 , doi= on Pith
  23. [43] Equipping Agents for the Real World with Agent Skills , year= on Pith
  24. [44] Self-Refine: Iterative Refinement with Self-Feedback , author=. 2023 , eprint= on Pith
  25. [45] Large Language Models as Optimizers , author=. 2024 , eprint= on Pith
  26. [46] Agrawal, Lakshya A and Tan, Shangyin and Soylu, Dilara and Ziems, Noah and Khare, Rishi and Opsahl-Ong, Krista and Singhvi, Arnav and Shandilya, Herumb and Ryan, Michael J and Jiang, Meng and Potts, Christopher and Sen, Koushik and Dimakis, Alexandros G. and Stoica, Ion and Kl... on PithPith's reviewarXiv
  27. [47] Du, Yiming and Huang, Wenyu and Zheng, Danna and Wang, Zhaowei and Montella, Sebastien and Lapata, Mirella and Wong, Kam-Fai and Pan, Jeff Z. , year=. Rethinking Memory in. 2505.00675 , archivePrefix= on PitharXiv
  28. [48] A Survey on the Memory Mechanism of Large Language Model-Based Agents , author=. ACM Transactions on Information Systems , volume=. 2025 , doi= on Pith
  29. [49] ReasoningBank: Scaling Agent Self-Evolving with Reasoning Memory , author=. The Fourteenth International Conference on Learning Representations , year= on Pith
  30. [50] MemRL: Self-Evolving Agents via Runtime Reinforcement Learning on Episodic Memory , author=. 2026 , eprint= on Pith
  31. [51] HaluMem: Evaluating Hallucinations in Memory Systems of Agents , author=. 2025 , eprint= on Pith
  32. [52] MemGPT: Towards LLMs as Operating Systems , author=. 2023 , eprint= on Pith
  33. [53] Zep: A Temporal Knowledge Graph Architecture for Agent Memory , author=. 2025 , eprint= on Pith
  34. [54] HiAgent: Hierarchical Working Memory Management for Solving Long-Horizon Agent Tasks with Large Language Model , author=. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=. 2025 , doi= on Pith
  35. [55] MemGen: Weaving Generative Latent Memory for Self-Evolving Agents , author=. 2025 , eprint= on Pith
  36. [56] MUSE-Autoskill: Self-Evolving Agents via Skill Creation, Memory, Management, and Evaluation , author=. 2026 , eprint= on Pith
  37. [57] SkillOS: Learning Skill Curation for Self-Evolving Agents , author=. 2026 , eprint= on Pith
  38. [58] AutoSkill: Experience-Driven Lifelong Learning via Skill Self-Evolution , author=. 2026 , eprint= on Pith
  39. [59] SkillHone: A Harness for Continual Agent Skill Evolution Through Persistent Decision History , author=. 2026 , eprint= on Pith
  40. [60] DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models , author=. 2025 , eprint= on Pith
  41. [61] Kimi K2: Open Agentic Intelligence , author=. 2025 , eprint= on Pith
  42. [62] Hello GPT-4o , author=. 2024 , url= on Pith
  43. [63] GPT-4o Mini: Advancing Cost-Efficient Intelligence , author=. 2024 , url= on Pith
  44. [64] Sentence-BERT: Sentence Embeddings Using Siamese BERT-Networks , author=. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing , pages=. 2019 , doi= on Pith
  45. [65] MiniLM: Deep Self-Attention Distillation for Task-Agnostic Compression of Pre-Trained Transformers , author=. Advances in Neural Information Processing Systems , volume=. 2020 , url= on Pith
  46. [66] Mohit Shridhar and Xingdi Yuan and Marc-Alexandre C\^ot\'e and Yonatan Bisk and Adam Trischler and Matthew Hausknecht , booktitle=. 2021 , url= on Pith
  47. [67] Wang, Yuyao and Zhang, Zhongjian and Chi, Mo and Yu, Kaichi and Li, Yuhan and Peng, Miao and Tong, Bing and Zhang, Chen and Zhou, Yan and Li, Jia , keywords =. 2026 , copyright =. doi:10.48550/ARXIV.2605.18421 , url = on PithPith's reviewarXivdoi:10.48550/arxiv.2605.18421
  48. [68] Ren, Tao and Luo, Weiyao and Yang, Hui and Zhu, Rongzhi and Huang, Xiang and Wu, Yuchuan and Chou, Bingxue and Ye, Jieping and Liang, Jiafeng and Li, Yongbin and Peng, Yijie , keywords =. Scaling Self-Evolving Agents via Parametric Memory , publisher =. 2026 , copyright =. doi... on PithPith's reviewarXivdoi:10.48550/arxiv.2606.04536
  49. [69] Bogdanov, Igor and Lung, Chung-Horng and Kunz, Thomas and Gao, Jie and Taylor, Adrian and Zaman, Marzia , year =. doi:10.1145/3786335.3813155 , booktitle = on Pithdoi:10.1145/3786335.3813155
  50. [70] Wei, Tianxin and Sachdeva, Noveen and Coleman, Benjamin and He, Zhankui and Bei, Yuanchen and Ning, Xuying and Ai, Mengting and Li, Yunzhe and He, Jingrui and Chi, Ed H. and Wang, Chi and Chen, Shuo and Pereira, Fernando and Kang, Wang-Cheng and Cheng, Derek Zhiyuan , keywords... on PithPith's reviewarXivdoi:10.48550/arxiv.2511.20857
  51. [71] 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 , year =. From Storage to Experience: A Survey on the Evolution of. doi:10.18653/v1/2026.findings-acl.2069 , booktitle = on Pithdoi:10.18653/v1/2026.findings-acl.2069
  52. [72] Wang, Xiaoxing and Liao, Ning and Wei, Shikun and Tang, Chen and Xiong, Feiyu , keywords =. 2026 , copyright =. doi:10.48550/ARXIV.2603.09716 , url = on Pithdoi:10.48550/arxiv.2603.09716
  53. [73] Yang, John and Jimenez, Carlos E. and Wettig, Alexander and Lieret, Kilian and Yao, Shunyu and Narasimhan, Karthik R. and Press, Ofir , booktitle =. 2024 , url = on Pith
  54. [74] Jimenez, Carlos E. and Yang, John and Wettig, Alexander and Yao, Shunyu and Pei, Kexin and Press, Ofir and Narasimhan, Karthik R. , booktitle =. 2024 , url = on Pith
Cited by 1 Pith paper

Reviewed papers on Pith that cite this one. The full citation record

  1. Discovering Effective Agent Memory Designs through Program Evolution cs.CL · Nov 2026

Challenges and discussion

Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.

No comments yet.

Sign in to comment