pith. sign in

arxiv: 2607.00394 · v1 · pith:7ACTY54Snew · submitted 2026-07-01 · 💻 cs.DB · cs.CL

When Classic Cache Policies Fail: Learning-Augmented Replacement for Semantic Retrieval Buffers

Yushi Sun , Bowen Cao , Wai Lam This is my paper

Pith reviewed 2026-07-02 04:58 UTC · model grok-4.3

classification 💻 cs.DB cs.CL
keywords semantic cache replacementlearning-augmented algorithmsonline competitive analysisregret boundsLLM retrieval bufferscache management policiesswitching costs
0
0 comments X

The pith

SOLAR achieves a constant competitive ratio of at most 3 for semantic cache replacement, independent of size and horizon, where FIFO scales as Omega(K).

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

The paper formalizes semantic cache replacement for LLM agent retrieval buffers as an online problem that includes switching costs and continuous hit quality based on embedding similarity. Experiments on MemoryBench datasets show that classic policies such as LRU and LFU underperform even the naive FIFO baseline because semantic workloads lack temporal locality and frequency concentration. SOLAR combines regret accumulation to time cache modifications with Bayesian online learning to select content from implicit feedback. It proves the policy maintains a competitive ratio bounded by 3 regardless of cache size or time horizon while its eviction regret nearly matches the known lower bound. The approach yields measured gains of 5 to 75 percent over FIFO at small capacities and reveals a phase transition at the working-set boundary.

Core claim

The paper establishes that the learning-augmented SOLAR policy achieves a constant competitive ratio of at most 3 independent of cache size K and horizon T, while FIFO has competitive ratio Omega(K), together with eviction regret O(sqrt(KT log T)) that matches the Omega(sqrt(KT)) lower bound up to logarithmic factors.

What carries the argument

SOLAR, which sets modification timing via regret accumulation (approximately 17 percent rate) and content selection via Bayesian online learning over implicit retrieval feedback.

If this is right

  • SOLAR maintains bounded performance independent of cache size K while FIFO degrades linearly with K.
  • Eviction regret scales as O(sqrt(KT log T)) and nearly meets the information-theoretic lower bound.
  • Relative improvement over FIFO reaches 5-75 percent when cache size is tight relative to the working set.
  • An inverted-U curve appears between retrieval pool size and quality, indicating capacity constraints arise from noise.
  • A clear phase transition occurs exactly at the working-set boundary.

Where Pith is reading between the lines

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

  • The constant-ratio result suggests that embedding-based continuous matching can replace binary hit/miss models in other online caching settings.
  • The inverted-U pool-size finding implies that retrieval quality may improve by deliberately limiting pool size rather than expanding storage.
  • Bayesian selection over implicit feedback could transfer to other online selection problems that lack explicit reward signals.
  • The regret-driven modification schedule may generalize to other online problems that combine exploration cost with continuous value.

Load-bearing premise

The online formalization with continuous hit quality and switching costs accurately models the behavior of real LLM agent retrieval workloads.

What would settle it

A workload constructed from the same embedding similarity model in which SOLAR's observed competitive ratio exceeds 3 or its regret exceeds O(sqrt(KT log T)) by more than logarithmic factors.

Figures

Figures reproduced from arXiv: 2607.00394 by Bowen Cao, Wai Lam, Yushi Sun.

Figure 3
Figure 3. Figure 3 [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figure 1
Figure 1. Figure 1: Average F1 by method and cache size on LoCoMo (one panel per [PITH_FULL_IMAGE:figures/full_fig_p011_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Learning dynamics: F1 vs. evaluation step on Lo [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Hit rate on cycling workloads (K=10) as a function [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Retrieval noise inverted-U curve: hit rate vs. pool [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗
read the original abstract

LLM agents increasingly rely on retrieval buffers to store and reuse past experience, yet the cache management policies governing these buffers remain largely ad-hoc. We formalize this as an online semantic cache replacement problem with switching costs, where items are matched by embedding similarity and hit quality is continuous rather than binary. Through experiments on two datasets from MemoryBench-Full (LoCoMo, DialSim) with 8 replacement policies, we reveal a surprising finding: classic heuristics (LRU, LFU) \emph{consistently underperform} the naive FIFO baseline on semantic workloads, due to the absence of temporal locality and frequency concentration. We propose SOLAR, a learning-augmented framework that derives modification timing from regret accumulation (achieving $\sim$17\% modification rate) and content selection from Bayesian online learning over implicit retrieval feedback. We prove SOLAR achieves a constant competitive ratio $\leq 3$, independent of cache size and horizon (vs.\ $\Omega(K)$ for FIFO), and eviction regret $O(\sqrt{KT\log T})$, matching the $\Omega(\sqrt{KT})$ lower bound up to logarithmic factors. Experiments demonstrate 5--75\% relative improvement over FIFO at tight cache sizes, with a clearly characterized phase transition at the working set boundary. Synthetic experiments with 5000-item pools further reveal an inverted-U relationship between pool size and retrieval quality, justifying capacity constraints as a retrieval noise phenomenon rather than a storage limitation.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper formalizes semantic cache replacement for LLM agent retrieval buffers as an online problem with continuous hit quality (via embedding similarity) and switching costs. It experimentally demonstrates that classic policies (LRU, LFU) underperform FIFO on LoCoMo and DialSim workloads from MemoryBench-Full due to absent temporal locality. It introduces the SOLAR framework, which uses regret accumulation to decide modification timing (reported at ~17% rate) and Bayesian online learning over implicit feedback for content selection. The central theoretical claims are a constant competitive ratio ≤3 (independent of cache size K and horizon T, versus Ω(K) for FIFO) together with eviction regret O(√(KT log T)) that matches the Ω(√(KT)) lower bound up to log factors. Experiments report 5--75% relative gains over FIFO at tight sizes, a phase transition at the working-set boundary, and an inverted-U relationship between pool size and quality in synthetic 5000-item tests.

Significance. If the stated competitive ratio and regret bounds hold under the continuous hit-quality model, the work would supply the first constant-competitive learning-augmented policy for semantic caches, together with a clear explanation for the failure of classic heuristics and reproducible experimental gains on standard MemoryBench datasets. The matching of the regret lower bound and the characterization of capacity constraints as a retrieval-noise phenomenon would be notable strengths for the DB and online-algorithms communities.

major comments (2)
  1. [Abstract] Abstract and theoretical analysis: the claim of a K- and T-independent competitive ratio ≤3 is presented without derivation steps or the specific inequalities on the continuous hit-quality function and switching-cost model that are required for the bound; it is therefore impossible to confirm that the LoCoMo/DialSim embeddings satisfy the Lipschitz or sublinear-regret conditions needed for the analysis to apply rather than producing high-variance scores that would invalidate the ratio.
  2. [Experiments] Experiments section: the ~17% modification rate is reported from the same experimental runs used to claim performance gains, so the timing parameter of SOLAR is fitted to the data rather than derived independently from the regret-accumulation analysis; this creates a circular dependence that undermines the claim that the framework delivers the stated bounds in a parameter-light manner.
minor comments (2)
  1. [Abstract] The abstract states that classic policies 'consistently underperform' FIFO but provides no per-dataset breakdown or statistical significance tests for the 8 policies evaluated.
  2. Notation for hit quality and switching costs is introduced without an explicit equation relating embedding similarity to the continuous quality value used in the regret analysis.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their thorough review and valuable feedback. We are pleased that the significance of the work is recognized and will address the concerns regarding the presentation of the theoretical results and the experimental methodology in a revised version of the manuscript.

read point-by-point responses

  1. Referee: [Abstract] Abstract and theoretical analysis: the claim of a K- and T-independent competitive ratio ≤3 is presented without derivation steps or the specific inequalities on the continuous hit-quality function and switching-cost model that are required for the bound; it is therefore impossible to confirm that the LoCoMo/DialSim embeddings satisfy the Lipschitz or sublinear-regret conditions needed for the analysis to apply rather than producing high-variance scores that would invalidate the ratio.

    Authors: The full derivation of the competitive ratio bound is provided in Section 4 of the manuscript, where we specify the required conditions on the hit-quality function (Lipschitz continuity with constant L) and the switching cost model (additive per eviction). The proof shows that under these conditions, the ratio is at most 3 independent of K and T by relating the online cost to the offline optimum via regret accumulation. Regarding the embeddings, we will include in the revision an empirical verification that the similarity scores from the LoCoMo and DialSim datasets satisfy the Lipschitz condition (estimated L=0.85) and do not exhibit high variance that would violate the sublinear regret assumption. This will make the applicability clear. revision: yes

  2. Referee: [Experiments] Experiments section: the ~17% modification rate is reported from the same experimental runs used to claim performance gains, so the timing parameter of SOLAR is fitted to the data rather than derived independently from the regret-accumulation analysis; this creates a circular dependence that undermines the claim that the framework delivers the stated bounds in a parameter-light manner.

    Authors: We clarify that the regret accumulation threshold used to determine modification timing is set based on the theoretical analysis to guarantee the O(√(KT log T)) regret bound, resulting in the observed ~17% rate as an outcome rather than an input. The parameter is not tuned on the performance data. To strengthen this, we will revise the experiments section to explicitly derive the threshold from the regret bound and include sensitivity analysis showing that nearby thresholds yield similar gains, confirming the parameter-light nature. revision: yes

Circularity Check

0 steps flagged

No circularity: theoretical bounds derived from online formulation independent of experimental fits

full rationale

The abstract states a proof of competitive ratio ≤3 and regret O(√(KT log T)) based on the formalized online problem with continuous hit quality and switching costs. The ~17% modification rate is reported from experiments as an observed outcome of the regret accumulation mechanism, not as an input parameter to which the proof is fitted or reduced. No equations, self-citations, or ansatzes in the provided text show the bounds reducing by construction to experimental data or prior self-work. The derivation chain remains self-contained against the stated model assumptions.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 1 invented entities

The central claims rest on the problem formalization, the regret accumulation rule, and Bayesian online learning over implicit feedback; the 17% rate is extracted from experimental runs rather than derived from first principles.

free parameters (1)
  • modification rate = ~17%
    Reported as ~17% from regret accumulation on the evaluated datasets; used to control when SOLAR updates the buffer.
axioms (2)
  • domain assumption Semantic retrieval workloads exhibit no temporal locality or frequency concentration sufficient for classic policies to succeed.
    Invoked to explain why LRU and LFU fail relative to FIFO.
  • domain assumption Continuous hit quality and switching costs can be modeled in an online learning setting that admits constant competitive ratio analysis.
    Required for the proof that the ratio is independent of cache size and horizon.
invented entities (1)
  • SOLAR framework no independent evidence
    purpose: Combines regret-based modification timing with Bayesian content selection for semantic cache replacement.
    New policy introduced to achieve the stated competitive ratio and empirical gains.

pith-pipeline@v0.9.1-grok · 5790 in / 1496 out tokens · 26059 ms · 2026-07-02T04:58:52.918495+00:00 · methodology

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

39 extracted references · 25 canonical work pages · 9 internal anchors

  1. [1]

    Shipra Agrawal and Navin Goyal. 2012. Analysis of Thompson Sampling for the Multi-armed Bandit Problem. InProceedings of the 25th Annual Conference on Learning Theory (Proceedings of Machine Learning Research), Shie Mannor, Nathan Srebro, and Robert C. Williamson (Eds.), Vol. 23. PMLR, Edinburgh, Scotland, 39.1–39.26. https://proceedings.mlr.press/v23/agr...

    2012

  2. [2]

    Qingyao Ai, Yichen Tang, Changyue Wang, Jianming Long, Weihang Su, and Yiqun Liu. 2025. MemoryBench: A Benchmark for Memory and Continual Learn- ing in LLM Systems.arXiv preprint arXiv:2510.17281(2025). arXiv:2510.17281

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  3. [3]

    Fu Bang. 2023. GPTCache: An Open-Source Semantic Cache for LLM Applica- tions. InProceedings of the 3rd Workshop for Natural Language Processing Open Source Software (NLP-OSS). 212–218. https://doi.org/10.18653/v1/2023.nlposs- 1.24

    work page doi:10.18653/v1/2023.nlposs- 2023

  4. [4]

    László A. Bélády. 1966. A Study of Replacement Algorithms for a Virtual-Storage Computer.IBM Systems Journal5, 2 (1966), 78–101. https://doi.org/10.1147/sj. 52.0078

    work page doi:10.1147/sj 1966

  5. [5]

    2005.Online Computation and Competitive Analysis

    Allan Borodin and Ran El-Yaniv. 2005.Online Computation and Competitive Analysis. Cambridge University Press

    2005

  6. [6]

    2022.LangChain

    Harrison Chase. 2022.LangChain. https://github.com/langchain-ai/langchain Accessed: 2026-06-16

    2022

  7. [7]

    Peng Chen, Hailiang Zhao, Jiaji Zhang, Xueyan Tang, Yixuan Wang, and Shuiguang Deng. 2025. Robustifying Learning-Augmented Caching Efficiently without Compromising 1-Consistency. InAdvances in Neural Information Pro- cessing Systems (NeurIPS). https://neurips.cc/virtual/2025/loc/san-diego/poster/ 116615

    2025

  8. [8]

    Prateek Chhikara et al . 2025. Mem0: Building Production-Ready AI Agents with Scalable Long-Term Memory.arXiv preprint arXiv:2504.19413(2025). arXiv:2504.19413

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  9. [9]

    Franklin, Björn Þór Jónsson, Divesh Srivastava, and Michael Tan

    Shaul Dar, Michael J. Franklin, Björn Þór Jónsson, Divesh Srivastava, and Michael Tan. 1996. Semantic Data Caching and Replacement. InProceedings of the 22nd International Conference on Very Large Data Bases (VLDB). 330–341. https: //www.vldb.org/conf/1996/P330.PDF

    1996

  10. [10]

    Gil Einziger, Roy Friedman, and Ben Manes. 2017. TinyLFU: A Highly Efficient Cache Admission Policy.ACM Transactions on Storage13, 4 (2017), 1–31. https: //doi.org/10.1145/3149371

    work page doi:10.1145/3149371 2017

  11. [11]

    Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, Meng Wang, and Haofen Wang. 2024. Retrieval-Augmented Gen- eration for Large Language Models: A Survey.arXiv preprint arXiv:2312.10997 (2024). arXiv:2312.10997

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  12. [12]

    Song Jiang and Xiaodong Zhang. 2002. LIRS: An Efficient Low Inter-reference Recency Set Replacement Policy to Improve Buffer Cache Performance. InPro- ceedings of the ACM SIGMETRICS International Conference on Measurement and Modeling of Computer Systems. 31–42. https://doi.org/10.1145/511334.511340

    work page doi:10.1145/511334.511340 2002

  13. [13]

    Theodore Johnson and Dennis Shasha. 1994. 2Q: A Low Overhead High Performance Buffer Management Replacement Algorithm. InProceedings of the 20th International Conference on Very Large Data Bases (VLDB). 439–450. https://www.vldb.org/conf/1994/P439.PDF

    1994

  14. [14]

    Jiho Kim, Woosog Chay, Hyeonji Hwang, Daeun Kyung, Hyunseung Chung, Eunbyeol Cho, Yohan Jo, and Edward Choi. 2024. DialSim: A Real-Time Simulator for Evaluating Long-Term Dialogue Understanding of Conversational Agents. arXiv preprint arXiv:2406.13144(2024). arXiv:2406.13144

    work page arXiv 2024

  15. [15]

    2020.Bandit Algorithms

    Tor Lattimore and Csaba Szepesvári. 2020.Bandit Algorithms. Cambridge University Press. https://doi.org/10.1017/9781108571401

    work page doi:10.1017/9781108571401 2020

  16. [16]

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, and Douwe Kiela. 2020. Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. InAdvances in Neural Informa- tion Processing Systems, Vol. 33. 9459–9474. arXiv:2005.11401

    work page internal anchor Pith review Pith/arXiv arXiv 2020

  17. [17]

    Evan Zheran Liu, Milad Hashemi, Kevin Swersky, Parthasarathy Ranganathan, and Junwhan Ahn. 2020. An Imitation Learning Approach for Cache Replacement. InProceedings of the 37th International Conference on Machine Learning (ICML). 6237–6247. arXiv:2006.16239

    work page arXiv 2020

  18. [18]

    Thodoris Lykouris and Sergei Vassilvitskii. 2018. Competitive Caching with Machine Learned Advice. InProceedings of the 35th International Conference on Machine Learning (ICML). 3296–3305. arXiv:1802.05399

    work page arXiv 2018

  19. [19]

    Adyasha Maharana, Dong-Ho Lee, Sergey Tuber, and Mohit Bansal. 2024. Evalu- ating Very Long-Term Conversational Memory of LLM Agents. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (ACL). https://doi.org/10.18653/v1/2024.acl-long.747

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

  20. [20]

    Nimrod Megiddo and Dharmendra S. Modha. 2003. ARC: A Self-Tuning, Low Overhead Replacement Cache. InProceedings of the 2nd USENIX Conference on File and Storage Technologies (FAST). 115–130. https://www.usenix.org/legacy/ events/fast03/tech/megiddo.html

    2003

  21. [21]

    O’Neil, Patrick E

    Elizabeth J. O’Neil, Patrick E. O’Neil, and Gerhard Weikum. 1993. The LRU-K Page Replacement Algorithm for Database Disk Buffering. InProceedings of the ACM SIGMOD International Conference on Management of Data. 297–306. https://doi.org/10.1145/170035.170081

    work page doi:10.1145/170035.170081 1993

  22. [22]

    MemGPT: Towards LLMs as Operating Systems

    Charles Packer, Vivian Fang, Shishir G. Patil, Kevin Lin, Sarah Wooders, and Joseph E. Gonzalez. 2023. MemGPT: Towards LLMs as Operating Systems.arXiv preprint arXiv:2310.08560(2023). arXiv:2310.08560

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  23. [23]

    O’Brien, Carrie J

    Joon Sung Park, Joseph C. O’Brien, Carrie J. Cai, Meredith Ringel Morris, Percy Liang, and Michael S. Bernstein. 2023. Generative Agents: Interactive Simulacra of Human Behavior. InProceedings of the 36th Annual ACM Symposium on User Interface Software and Technology (UIST). 1–22. https://doi.org/10.1145/3586183. 3606763

    work page doi:10.1145/3586183 2023

  24. [24]

    Dhruv Rohatgi. 2020. Near-Optimal Bounds for Online Caching with Machine Learned Advice. InProceedings of the 31st Annual ACM-SIAM Symposium on Discrete Algorithms (SODA). 1834–1845. arXiv:1910.12172

    work page arXiv 2020

  25. [25]

    Sleator and Robert E

    Daniel D. Sleator and Robert E. Tarjan. 1985. Amortized Efficiency of List Update and Paging Rules.Commun. ACM28, 2 (1985), 202–208. https://doi.org/10.1145/ 2786.2793

    work page arXiv 1985

  26. [26]

    Berger, Kai Li, and Wyatt Lloyd

    Zhenyu Song, Daniel S. Berger, Kai Li, and Wyatt Lloyd. 2020. Learning Relaxed Belady for Content Distribution Network Caching. InProceedings of the 17th USENIX Symposium on Networked Systems Design and Implementation (NSDI). 529–544. https://www.usenix.org/conference/nsdi20/presentation/song

    2020

  27. [27]

    Changzhi Sun, Xiangyu Chen, Jixiang Luo, Dell Zhang, and Xuelong Li. 2025. Beyond Heuristics: A Decision-Theoretic Framework for Agent Memory Man- agement.arXiv preprint arXiv:2512.21567(2025). arXiv:2512.21567

    work page arXiv 2025

  28. [28]

    Yushi Sun, Bowen Cao, Dong Fang, Lingfeng Su, and Wai Lam. 2026. GRAVITY: Architecture-Agnostic Structured Anchoring for Long-Horizon Conversational Memory.arXiv preprint(2026)

    2026

  29. [29]

    Yushi Sun, Bowen Cao, and Wai Lam. 2026. SOLAR: Complete Proofs (Extended Supplement). https://github.com/ysunbp/SOLAR/blob/main/proof.pdf. Accessed: 2026

    2026

  30. [30]

    Yushi Sun and Lei Chen. 2026. CacheRAG: A Semantic Caching System for Retrieval-Augmented Generation in Knowledge Graph Question Answering. arXiv preprint(2026). arXiv:2604.26176

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  31. [31]

    Yushi Sun, Kai Sun, Ethan Yifan Xu, Xiao Yang, Xin Luna Dong, Nan Tang, and Lei Chen. 2025. KERAG: Knowledge-Enhanced Retrieval-Augmented Generation for Advanced Question Answering. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing (EMNLP)

    2025

  32. [32]

    Thompson

    William R. Thompson. 1933. On the Likelihood that One Unknown Probability Exceeds Another in View of the Evidence of Two Samples.Biometrika25, 3/4 (1933), 285–294. https://doi.org/10.1093/biomet/25.3-4.285

    work page doi:10.1093/biomet/25.3-4.285 1933

  33. [33]

    Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. 2023. Voyager: An Open-Ended Embodied Agent with Large Language Models.arXiv preprint arXiv:2305.16291(2023). arXiv:2305.16291

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  34. [34]

    Kin-Yeung Wong. 2006. Web cache replacement policies: a pragmatic approach. IEEE Network20, 1 (2006), 28–34

    2006

  35. [35]

    Xiao Yang, Kai Sun, Hao Xin, Yushi Sun, Nikhil Bhalla, Xiangsen Chen, Sajal Choudhary, Rongze Daniel Gui, Ziran Will Jiang, Ziyu Jiang, Lingkun Kong, Brian Moran, Jiaqi Wang, Yifan Ethan Xu, An Yang, Eting Yuan, Hanwen Zha, Nan Tang, Lei Chen, Nicolas Scheffer, Yue Liu, Nirav Shah, Rakesh Wanga, Anuj Kumar, Wen-tau Yih, and Xin Luna Dong. 2024. CRAG – Com...

    2024

  36. [36]

    Guilin Zhang, Wei Jiang, Xiejiashan Wang, Aisha Behr, Kai Zhao, Jeffrey Fried- man, Xu Chu, and Amine Anoun. 2026. Adaptive Memory Admission Control for LLM Agents.arXiv preprint arXiv:2603.04549(2026). arXiv:2603.04549

    work page arXiv 2026

  37. [37]

    Yuxiang Zhang, Jiangming Shu, Ye Ma, Xueyuan Lin, Shangxi Wu, and Jitao Sang. 2025. Memory as Action: Autonomous Context Curation for Long-Horizon Agentic Tasks.arXiv preprint arXiv:2510.12635(2025). arXiv:2510.12635

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  38. [38]

    Zeyu Zhang et al. 2024. A Survey on the Memory Mechanism of Large Language Model Based Agents.arXiv preprint arXiv:2404.13501(2024). arXiv:2404.13501

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  39. [39]

    Wanjun Zhong, Lianghong Guo, Qiqi Gao, He Ye, and Yanlin Wang. 2024. Mem- oryBank: Enhancing Large Language Models with Long-Term Memory. InPro- ceedings of the AAAI Conference on Artificial Intelligence, Vol. 38. 19724–19731. https://ojs.aaai.org/index.php/AAAI/article/view/29946 13

    2024