Pith.

Artificial Intelligence · arXiv:2608.13883 v1 · submitted 14 Aug 2026 · 15 references

MemoryLake on MemoryArena: A Matched Study of Agent Memory Backends

Chaoqun Zhan, Qiang Zhou, Guannan Li, Zhenqiang Huang, Qianjin Wang

The paper says

“Most agent-memory benchmarks test post-hoc recall, whereas MemoryArena evaluates whether memory supports interdependent, multi-session task completion. We compare MemoryLake, a structured multi-track memory backend, with Mem0, text-embedding-3-small vector RAG, and a long-context control across all five MemoryArena domains.”

more of the abstract

The systems share the same agent framework, requested gpt-5-mini model alias, task samples, and scoring code; the memory integration is the intentionally changed component. Because each backend bundles write, retrieval, consolidation, budgeting, and prompt-assembly choices, the study is a matched system-level comparison, not a representation-only ablation or a cost-matched experiment. On the shared evaluation sets, MemoryLake has the highest observed success rate (SR) in mathematics (9/40), physics (12/20), and progressive retrieval (4/20). Every system has zero SR in travel planning, and web shopping yields a single bundle-level success (long context, 1/150); MemoryLake ranks third on both the travel soft process score and shopping step match. Following MemoryArena's suite-level convention, a post-hoc equal-weight average over the five SRs is 20.5% for MemoryLake versus 13.6% for the best comparator. These are point estimates: sample sizes are modest, confidence intervals overlap, and we do not report paired significance tests. A separate MemoryLake-only run over all 221 progressive queries yields a failure-counted SR of 26.7% (59/221) and is not a baseline comparison. The results support a workload-dependent view of memory backends and an observed lead among the four evaluated systems on the shared sets; they do not establish benchmark-wide state of the art or a causal advantage of representation structure.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 27 Aug 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. An unusually honest benchmark comparison from a commercial team: the point-estimate ranking is credible and well-scoped, but the 'matched' premise rests on an unverifiable model-equivalence claim that should be audited before the ranking is used. This line is wrong

3 major objections · 4 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 15 references say.

The paper says, at Sections 3.1, 4, and 6.2

The review did not quote this passage.

Pith says: Objection 1 of 3 · major

The matched-comparison premise is not externally verifiable. Section 4 states that all LLM operations in MemoryLake, including internal consolidation, extraction, and assembly, use the same requested gpt-5-mini alias and that no larger or auxiliary model is used, but Section 6.2 concedes that resolved snapshot IDs, complete run dates, and all decoding and retry settings are not available in the public artifact.

more of objection 1

Because MemoryLake's backend is proprietary, a reader cannot check whether the internal calls that determine what enters the agent's context used a different snapshot, different sampling parameters, or an auxiliary model. If those calls used a stronger configuration, the observed lead in mathematics, physics, and progressive retrieval could reflect model strength rather than memory design. Please provide a call audit with resolved model/snapshot IDs and decoding settings for all systems, or explicitly narrow all comparative language to 'under the self-reported alias configuration' and remove the implication that model equivalence was verified.

The paper says, at Sections 3.4, 5.1, and 6.2

The review did not quote this passage.

Pith says: Objection 2 of 3 · major

The LLM-judge dependence is acknowledged but not quantified, and it is load-bearing because the ranking margins are very small. Mathematics, physics, and progressive retrieval all rely on the benchmark's LLM-based judging, and the paper reports no inter-judge agreement, no independent re-judging, and no blind audit.

more of objection 2

The decisive differences are one or two instances in mathematics (9/40 vs. 8/40) and progressive retrieval (4/20 vs. 3/20 or 2/20), so even modest judge instability could change a domain-level rank. The public artifact stores parsed labels but not complete judge transcripts. Please add judge-stability evidence, such as inter-judge agreement on a sample, or a blind re-audit of all ranking-critical instances, and report the sensitivity of the observed ranking to label perturbations.

The paper says, at Sections 3.1, 5.3, and 6.2 (Tables 3 and 4)

The review did not quote this passage.

Pith says: Objection 3 of 3 · major

The Mem0 capped re-run means that one comparator row is not measured under the same code as the other systems. Section 6.2 states that the per-write size cap is gated on the backend name and that no other system received it, and Table 3 marks the row accordingly.

more of objection 3

This disclosure is adequate, but the paper still calls the study 'matched' and uses the capped row in the main progressive comparison and in the macro-average. The uncapped run, which scores 0/20 for progressive SR, would actually strengthen MemoryLake's progressive lead, so this issue does not favor the authors, but it does undermine the like-for-like comparison. Please make the primary analysis either exclude the non-identical Mem0 row from cross-system ranking or present the uncapped run as primary with the capped run as a sensitivity analysis, and adjust the 'matched' wording accordingly.

4 minor objections

The paper says, at Abstract and body text

Pith says: minor objection 1 of 4

There are several spacing/typographical artifacts in the extracted text, such as 'Mostagent-memorybenchmarkstestpost-hocrecall' in the abstract; these should be cleaned in the final version.

The paper says, at Section 3.3

Pith says: minor objection 2 of 4

The progressive subset selection is described as 'selected before scoring' and then as a 'verified reconstruction' recovered from the published subset rather than the original script. Please clarify the timeline: if the reconstruction happened after scoring, state so explicitly and explain how the possibility of outcome-informed selection was excluded.

show 2 more

The paper says, at Section 6.2, Baseline coverage

Pith says: minor objection 3 of 4

The text names Letta, GraphRAG, and ReasoningBank as structured-memory systems but does not include citations for them; either add citations or mark them as systems not included in the reference list.

The paper says, at Section 5.3 and Table 7

Pith says: minor objection 4 of 4

The sentence about Vector RAG's PS denominator reads clearly, but the accompanying discussion of MemoryLake's coverage and assembly hypothesis would benefit from an explicit statement that no assembly ablation was run; the current text says this indirectly, and making it explicit in the main text would reduce the risk of over-reading.

What would settle it

Ask the authors to release a complete LLM-call audit log for all four systems: per call, the requested alias, resolved snapshot ID, model family, sampling parameters, retries, and timestamps. The check is to confirm that every MemoryLake-internal call (consolidation, extraction, assembly) resolved to the same gpt-5-mini snapshot family as the task-agent calls in all comparator runs, and that no call went to a larger or auxiliary model. If any MemoryLake-internal call used a different resolved snapshot or a different model family, the matched premise fails and the headline ranking would need to be re-interpreted as model-mixed rather than memory-backend-comparative. This is a single concrete verification step that the authors can perform from their own infrastructure logs.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 27 Aug 2026

  1. Reader holds up, with conditions confidence high
  2. Skeptic agrees 41 s · 12,166 tokens read, 4,078 written
  3. Editor wrote the verdict above 21 s · 12,931 tokens read, 2,063 written
  4. Referee major revision: 3 major objections, 4 minor 44 s · 12,486 tokens read, 4,648 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

If you work on agent memory, this paper is worth your time. It is a matched four-system comparison on MemoryArena across all five domains, and the team did the rare thing of reporting exact counts, Wilson intervals, negative results, and a post-hoc macro-average clearly labeled as post hoc. The arithmetic checks out: I re-derived the 20.5% versus 13.6% macro-average and the intervals match the stated counts. The progressive-retrieval aggregation script is a useful addition since the benchmark doesn't ship one, and the artifacts include per-instance scored outputs so the aggregates can be recomputed. Credit where due: this is a transparent evaluation from a commercial group, with explicit statements about what is and isn't claimed.

The soft spots are real but mostly disclosed. The load-bearing one is model equivalence. Section 4 says all MemoryLake internal LLM operations use the same requested gpt-5-mini alias, but the engine is proprietary, and Section 6.2 admits there are no resolved snapshot IDs, run dates, or decoding settings. A requested alias is not a pinned model. If the internal consolidation, extraction, or assembly calls resolved to a stronger snapshot or different sampling parameters, the observed lead in mathematics, physics, and progressive retrieval could reflect model strength rather than memory design. The stress-test note lands here. The paper flags this as a reproducibility limitation, which is honest, but it remains the one premise that the matched-comparison framing depends on. A resolved-snapshot audit or call log would settle it.

Other concerns are proportionate. LLM judge stability is not independently validated; the internal recomputation is an arithmetic check, not a fresh judging pass. Mem0's progressive row is a capped re-run, not a like-for-like measurement, though the unmodified run is also reported, which mitigates the concern. Sample sizes are modest and the key margins are one to three instances with overlapping intervals. All of this is disclosed, so the paper's claims are appropriately bounded.

Who is this for? Practitioners choosing memory backends and researchers designing agent-memory evaluations. The workload-dependence conclusion is plausible and the transparency is a model for how commercial evaluations should be written. It deserves a serious referee. I would send it to peer review, with the expectation that reviewers press for a model-equivalence audit or a clear falsification test of the matched premise.

The strongest objection

The study's central inference — that the observed success-rate differences reflect memory-backend configuration rather than model strength — rests on the assertion in Section 4 that 'All LLM operations in MemoryLake ... use the same requested gpt-5-mini alias; no larger or auxiliary language model is used inside the backend.' This is a self-report about a proprietary engine. Section 6.2 concedes that 'Resolved snapshot IDs, complete run dates, and all decoding and retry settings are not available in the current public artifact.' A requested alias is not the same as a pinned model: hosted aliases can resolve to different snapshots or use different decoding defaults, and the internal consolidation, extraction, and assembly calls are exactly the calls that determine what content enters the agent's context. If those calls used a stronger snapshot, different sampling parameters, or an auxiliary model, then the 'matched' comparison in Section 3.1 is not matched in the load-bearing way, and MemoryLake's lead in mathematics, physics, and progressive retrieval could reflect model behavior rather than memory design. This is not an internal inconsistency; the arithmetic and Wilson intervals check out, and the claim is explicitly a point estimate. But the paper's value as a matched study depends on this unverified premise, so the concern is load-bearing. A resolved-snapshot and call-audit log would directly test it.

The reader's weakest assumption identified exactly the same load-bearing premise: the requested gpt-5-mini alias is not verified to resolve to comparable model behavior inside the proprietary MemoryLake backend. I agree that this is the most load-bearing concern. The paper's headline claim is deliberately hedged as an observed point estimate, and the internal arithmetic is consistent, but the 'matched study' framing gives the result its meaning. If the model-equivalence premise is false, the observed SR ordering cannot be attributed to memory design, and the central comparison collapses into an uncontrolled model-strength comparison. The concern is not that the authors are dishonest; it is that the current public artifact cannot rule out the alternative. The paper itself flags this in Section 6.2, which is appropriate, but a self-flagged limitation does not settle the question. Other concerns — the non-random progressive subset, the Mem0 size cap, judge dependence, and the unmodified output-token budget — are real, but they are either disclosed with dual reporting or affect only one domain; the model-equivalence premise affects every domain and the entire matched-comparison interpretation. A resolved-snapshot audit is the single concrete test that would determine whether the premise holds. If the audit passes, the observed lead stands as a matched comparison under the stated protocol. If it fails, the observed numbers may still be accurate, but they would no longer support the paper's title claim of a matched study. Until then, CONDITIONAL is the right verdict, so I recommend no change to the reader's verdict.

The referee's report

If the results hold, the paper is a useful, unusually transparent empirical contribution to agent-memory evaluation in the MemoryArena setting. It ships per-instance scored outputs and aggregation scripts, reports exact denominators, and separates process metrics from end-to-end success. The workload-dependent ranking—MemoryLake leading in formal reasoning and progressive retrieval while long context leads travel and shopping process metrics—is a plausible and falsifiable finding. The main value, however, is bounded by the modest margins (one to three instances) and by three verification gaps: the matched-model premise is self-reported for a proprietary engine, the LLM judge stability is unquantified, and one comparator row was produced under a system-specific code change. These gaps are disclosed in the manuscript, which is commendable, but they are load-bearing for the central claim that the observed differences reflect memory-backend design rather than uncontrolled model, judge, or code differences.

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

The paper sets out to test whether a structured multi-track memory backend improves an agent's ability to complete interdependent, multi-session tasks, as opposed to merely recalling facts. It compares MemoryLake against three baselines—Mem0, embedding-based vector retrieval, and a verbatim long-context control—with the same agent framework, the same requested base model, the same task samples, and the same scoring code across all systems. On the shared evaluation sets, MemoryLake posts the highest observed success rate in mathematics (9/40), physics (12/20), and progressive retrieval (4/20), and its post-hoc equal-weight macro-average SR is 20.5% versus 13.6% for the best baseline. The authors deliberately frame these as point estimates: samples are small, confidence intervals overlap, and no paired significance tests are reported. The intended lesson is that memory backends are not uniformly ordered—structured memory helps where later decisions depend on exact conclusions or aggregated findings, while long context still wins where broad verbatim replay matters.

What carries the argument. The mechanism that carries the argument is MemoryLake's structured multi-track memory, which maintains three content tracks under different presence policies: confirmed conclusions are prioritized for inclusion in later prompts, supporting evidence is retrieved on demand, and reusable experience is consolidated for transfer across subtasks. Recall assembles these tracks under a bounded context budget, with fallback to an empty memory context if the internal service fails. The matched protocol—identical agent framework, requested model alias, task samples, and scoring scripts, with the memory integration as the intentionally changed component—is what allows the paper to attribute observed differences to the backend configuration rather than to the agent or evaluator. A supporting methodological piece is the paper's progressive-retrieval aggregation, which counts unanswered slots as incorrect instead of dividing by answered slots, changing measured process scores for systems with low slot coverage.

Core claim. The central discovery, on the paper's own terms, is a workload-dependent ranking among four memory configurations on MemoryArena. MemoryLake—a structured multi-track backend that separates confirmed conclusions, supporting evidence, and reusable experience under different presence policies—achieves the highest observed end-to-end success rate (SR) in mathematics (9/40), physics (12/20), and progressive retrieval (4/20), and the highest post-hoc macro-average SR (20.5%) against the best comparator (13.6%, long context). The paper explicitly bounds the claim: it is an observed lead among the four evaluated systems on the shared sets, not statistical significance, not benchmark-wide state of the art, and not a causal attribution to any single representation mechanism. Process metrics show no uniform lead—long context leads the travel soft process score and shopping step match, all systems score zero SR in travel, and shopping yields one bundle-level success—which is the evidence for the workload-dependent conclusion.

The premise it rests on. The whole comparison assumes that MemoryLake genuinely runs the same requested base model inside its proprietary backend as the other systems do; if a stronger model or a different snapshot was actually used, the observed ranking could reflect model power rather than memory architecture.

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

On tasks where later decisions depend on exact intermediate conclusions or aggregation of findings, a structured memory with prioritized conclusion presence can outperform both embedding-based retrieval and verbatim long context on end-to-end success.

Memory-backend rankings are workload-dependent: the same system that leads formal reasoning trails in travel planning and shopping, so single-number leaderboards can mislead about which backend suits a given application.

Reporting process metrics alongside final success matters; the paper shows a backend without the top process score (MemoryLake in progressive retrieval) can still have the top final success, and vice versa.

Because the paper's suite-level average mixes full-domain and subset runs, a stronger claim about benchmark-wide superiority would require full common-set runs, paired inference, and pinned model snapshots.

The paper's strongest comparative result rests on an unverifiable premise: that MemoryLake's proprietary engine really did use the same requested base-model alias internally. An independent replication that pins and logs the resolved model snapshot for every run would settle whether the observed lead comes from memory design or from a model-strength confound.

The paper's directional argument that the embedder mismatch favors its competitor—because the memory backend's embedder has strengths that are inactive in this English, short-chunk, dense-only benchmark—could be tested directly with an embedder-matched control run; such a test would either remove or confirm the main residual confound.

The workload-dependent pattern suggests a follow-up design: a backend that adapts its presence policy per task type (prioritized conclusions for formal reasoning, high-fidelity replay for travel) should dominate any fixed-policy backend on the same benchmark—a testable prediction the paper does not make.

The paper documents that unanswered progressive slots can vanish from the scorer rather than scoring zero (a large serialization object overflows the judge's context window), which implies published MemoryArena scores from other pipelines may overstate success rates if they rely on the default denominator; this is an implication for the field, not a claim of the paper.

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

Most agent-memory benchmarks test post-hoc recall, whereas MemoryArena evaluates whether memory supports interdependent, multi-session task completion. We compare MemoryLake, a structured multi-track memory backend, with Mem0, text-embedding-3-small vector RAG, and a long-context control across all five MemoryArena domains. The systems share the same agent framework, requested gpt-5-mini model alias, task samples, and scoring code; the memory integration is the intentionally changed component. Because each backend bundles write, retrieval, consolidation, budgeting, and prompt-assembly choices, the study is a matched system-level comparison, not a representation-only ablation or a cost-matched experiment. On the shared evaluation sets, MemoryLake has the highest observed success rate (SR) in mathematics (9/40), physics (12/20), and progressive retrieval (4/20). Every system has zero SR in travel planning, and web shopping yields a single bundle-level success (long context, 1/150); MemoryLake ranks third on both the travel soft process score and shopping step match. Following MemoryArena's suite-level convention, a post-hoc equal-weight average over the five SRs is 20.5% for MemoryLake versus 13.6% for the best comparator. These are point estimates: sample sizes are modest, confidence intervals overlap, and we do not report paired significance tests. A separate MemoryLake-only run over all 221 progressive queries yields a failure-counted SR of 26.7% (59/221) and is not a baseline comparison. The results support a workload-dependent view of memory backends and an observed lead among the four evaluated systems on the shared sets; they do not establish benchmark-wide state of the art or a causal advantage of representation structure.

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 · 2026-09-23

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

“The results support a workload-dependent view of memory backends and an observed lead among the four evaluated systems on the shared sets; they do not establish benchmark-wide state of the art or a causal advantage of representation structure.”

How the main claim is stated: hedged possibility 0.61

What the abstract itself states, as probabilities: claims first 0.05 · states limitation 0.98 · human participants 0.04 · controlled comparison 0.95 · randomized assignment 0.02 · states quantitative result 0.99 · compares with prior results 0.92 · mentions code or data availability 0.05

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
Most agent-memory benchmarks test post-hoc recall, whereas MemoryArena evaluates whether memory supports interdependent, multi-session task completion. gap 0.62 0.23 0.05 asserted 1.00
We compare MemoryLake, a structured multi-track memory backend, with Mem0, text-embedding-3-small vector RAG, and a long-context control across all five MemoryArena domains. method 0.54 0.92 0.03 asserted 0.99
The systems share the same agent framework, requested gpt-5-mini model alias, task samples, and scoring code; the memory integration is the intentionally changed component. method 1.00 0.87 0.03 asserted 1.00
Because each backend bundles write, retrieval, consolidation, budgeting, and prompt-assembly choices, the study is a matched system-level comparison, not a representation-only ablation or a cost-matched experiment. method 0.81 0.76 0.05 asserted 1.00
On the shared evaluation sets, MemoryLake has the highest observed success rate (SR) in mathematics (9/40), physics (12/20), and progressive retrieval (4/20). result 1.00 0.96 0.07 asserted 1.00
Every system has zero SR in travel planning, and web shopping yields a single bundle-level success (long context, 1/150); MemoryLake ranks third on both the travel soft process score and shopping step match. result 1.00 0.95 0.04 asserted 1.00
Following MemoryArena's suite-level convention, a post-hoc equal-weight average over the five SRs is 20.5% for MemoryLake versus 13.6% for the best comparator. result 0.98 0.94 0.12 asserted 1.00
These are point estimates: sample sizes are modest, confidence intervals overlap, and we do not report paired significance tests. limitation 1.00 0.92 0.04 asserted 0.98
A separate MemoryLake-only run over all 221 progressive queries yields a failure-counted SR of 26.7% (59/221) and is not a baseline comparison. result 0.96 0.97 0.03 asserted 1.00
The results support a workload-dependent view of memory backends and an observed lead among the four evaluated systems on the shared sets; they do not establish benchmark-wide state of the art or a causal advantage of representation structure. interpretation 1.00 0.88 0.91 hedged possibility 0.61

What kind of paper the classifier reports this is: empirical study 1.00 · evidence from observation or experiment 0.99

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, 2026-09-23: 1 request, 7,871 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

References (15)
  1. [1] A. Maharana, D.-H. Lee, S. Tulyakov, M. Bansal, F. Barbieri, and Y. Fang. Evaluating very long-term conversational memory of LLM agents. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics, 2024 on Pith
  2. [2] D. Wu, H. Wang, W. Yu, Y. Zhang, K.-W. Chang, and D. Yu. LongMemEval: Benchmark- ing chat assistants on long-term interactive memory. InInternational Conference on Learning Representations, 2025 on Pith
  3. [3] Y. Hu, Y. Wang, and J. McAuley. Evaluating memory in LLM agents via incremental multi- turn interactions.arXiv preprint arXiv:2507.05257, 2025. 13 on PithPith's reviewarXiv
  4. [4] Q. Ai, Y. Tang, C. Wang, J. Long, W. Su, and Y. Liu. MemoryBench: A benchmark for memory and continual learning in LLM systems.arXiv preprint arXiv:2510.17281, 2025 on PithPith's reviewarXiv
  5. [5] Z. He, Y. Wang, C. Zhi, Y. Hu, T.-P. Chen, L. Yin, Z. Chen, T. A. Wu, S. Ouyang, Z. Wang, J. Pei, J. McAuley, Y. Choi, and A. Pentland. MemoryArena: Benchmarking agent memory in interdependent multi-session agentic tasks.arXiv preprint arXiv:2602.16313, 2026 on PithPith's reviewarXiv
  6. [6] S. Zhou, F. F. Xu, H. Zhu, X. Zhou, R. Lo, A. Sridhar, X. Cheng, T. Ou, Y. Bisk, D. Fried, U. Alon, and G. Neubig. WebArena: A realistic web environment for building autonomous agents. InInternational Conference on Learning Representations, 2024 on Pith
  7. [7] S. Yao, H. Chen, J. Yang, and K. Narasimhan. WebShop: Towards scalable real-world web interaction with grounded language agents. InAdvances in Neural Information Processing Sys- tems, 2022 on Pith
  8. [8] C. E. Jimenez, J. Yang, A. Wettig, S. Yao, K. Pei, O. Press, and K. Narasimhan. SWE- bench: Can language models resolve real-world GitHub issues? InInternational Conference on Learning Representations, 2024 on Pith
  9. [9] P. Chhikara, D. Khant, S. Aryan, T. Singh, and D. Yadav. Mem0: Building production-ready AI agents with scalable long-term memory.arXiv preprint arXiv:2504.19413, 2025 on PithPith's reviewarXiv
  10. [10] C. Packer, S. Wooders, K. Lin, V. Fang, S. G. Patil, I. Stoica, and J. E. Gonzalez. MemGPT: Towards LLMs as operating systems.arXiv preprint arXiv:2310.08560, 2023 on PithPith's reviewarXiv
  11. [11] D. Edge, H. Trinh, N. Cheng, J. Bradley, A. Chao, A. Mody, S. Truitt, and J. Larson. From local to global: A graph RAG approach to query-focused summarization.arXiv preprint arXiv:2404.16130, 2024 on PithPith's reviewarXiv
  12. [12] H. Qian, Z. Liu, P. Zhang, K. Mao, D. Lian, Z. Dou, and T. Huang. MemoRAG: Boosting long context processing with global memory-enhanced retrieval augmentation. InProceedings of the ACM Web Conference, pages 2366–2377, 2025 on Pith
  13. [13] S. Ouyang, J. Yan, I.-H. Hsu, Y. Chen, K. Jiang, Z. Wang, R. Han, T. T. Le, S. Daruki, X. Tang, et al. ReasoningBank: Scaling agent self-evolving with reasoning memory.arXiv preprint arXiv:2509.25140, 2025 on PithPith's reviewarXiv
  14. [14] M. Zou, Z. Guo, L. Liang, Z. Wang, Q. Wang, Q. Wen, I. King, L. Qu, and Z. Xu. Remember the decision, not the description: A rate-distortion framework for agent memory.arXiv preprint arXiv:2605.10870, 2026 on PithPith's reviewarXiv
  15. [15] B. Tang, Y. Zhang, G. Zhuang, W. Wei, G. Zheng, L. Xie, Y. Tan, F. Xiong, Q. Yang, E. Chung, and Z. Li. From memory to skills: Evidence-grounded co-evolution governance for long-horizon LLM agents.arXiv preprint arXiv:2607.16621, 2026. A Alignment with the MemoryArena v1 Relea... on PithPith's reviewarXiv
How the literature uses it

What this paper does with the arXiv papers it cites

Each citing sentence is taken from the citing paper's LaTeX source and typed by Pith's classifier. A label appears at probability 0.8 or more (0.95 for support), a contest at 0.7 or more; the quoted words are the citing authors'. How each label was checked · JSON

What it cites · 5 sentences without a confident label

Uses the method or tool of 1 paper
Uses data or values from 1 paper
Cites as background 7 papers

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