Pith. sign in

REVIEW 4 major objections 5 minor 43 references

Agent error correction can be computed offline as a graph edit path, replacing test-time reflection loops.

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

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

T0 review · deepseek-v4-flash

2026-08-02 03:25 UTC pith:63KEFERJ

load-bearing objection Novel graph-edit-path memory for one-shot agent correction, but the empirical comparison is unfair until baselines get equal access to expert demonstrations. the 4 major comments →

arxiv 2607.13884 v1 pith:63KEFERJ submitted 2026-07-15 cs.AI

Experience Memory Graph: One-Shot Error Correction for Agents

Wenjun Wang , Yuchen Fang , Fengrui Liu , Zibo Liang , Kai Zheng This is my paper
classification cs.AI
keywords LLM agentsagent memoryexperience memorygraph edit pathcommon subgraphone-shot error correctionaction decision graphreflection-free inference
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper tries to establish that the failure-recovery loop of LLM agents—reflect, replay, retry—can be replaced by an offline, deterministic computation. It converts each training task's failed exploration trajectory and its paired successful expert trajectory into directed action decision graphs, matches those graphs, and extracts the shortest graph edit path: a concrete list of actions to keep, delete, or relabel under specific observations. These corrections, plus common subgraphs capturing successful workflows, are stored in a memory graph whose nodes hold task-specific insights and whose edges hold cross-task insights. At test time the agent retrieves the relevant insights and executes once. If the paper's results hold, iterative trial-and-error reflection becomes unnecessary for tasks that resemble training examples, cutting latency and API cost while improving success rates, especially for small models and long-horizon tasks.

Core claim

The central claim is that error correction is not inherently a reasoning problem that needs LLM self-reflection; it can be formulated as graph matching. By representing trajectories as graphs with action nodes and observation-labeled edges, and by matching failed graphs against expert graphs, one obtains a common subgraph (what was already right) and a graph edit path (what to add, delete, or relabel) that is a valid, executable repair plan. Stored offline in a memory graph and retrieved by query similarity at test time, this plan yields one-shot execution with no loops. The paper reports that EMG consistently outperforms state-of-the-art reflection baselines in success rate and average rewa

What carries the argument

The central object is the directed action decision graph: nodes are normalized action tuples (type, object, receptacle) with raw action text, edges are labeled by the observation preceding the action, a virtual INIT node anchors the first action, repeated actions reuse nodes, and consecutive invalid actions are parallelized from the same valid state so they share an observation label. Matching two such graphs uses a Fused Gromov-Wasserstein optimal transport solver with node and edge labels, followed by linear assignment to a hard node matching; from that matching, the paper's Algorithms 2 and 3 extract the common subgraph and the graph edit path. The graph edit path is the load-bearing arti

Load-bearing premise

The method assumes every training task has a paired successful expert trajectory to match against; without that gold trajectory, there is no target for the graph edit path, and the correction insight cannot be computed — a limitation the paper acknowledges and defers to future work.

What would settle it

On a suite of long-horizon tasks that include paired expert trajectories, if EMG's one-shot success rate fails to beat a 5-iteration reflection baseline given the same total API budget, the paper's central outperformance claim would be falsified.

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

If this is right

  • Test-time cost drops to a single execution: no repeated LLM reflection and no replanning loops, so latency and API spend are roughly the cost of one rollout.
  • Correction quality moves from prompt-sensitivity to structural comparison: the same correction is computed algorithmically, independent of how well the test-time model can introspect.
  • Failure recovery works for small models: the paper's results show the largest relative gains on a 4B-parameter model, suggesting structured memory can substitute for missing reasoning capacity.
  • Cross-task edges transfer knowledge to unseen tasks: matching expert graphs of similar queries yields common subgraphs that generalize beyond task-specific memory banks.
  • The offline construction means the memory graph, once built, can be reused across any test set with similar queries, amortizing the training cost.

Where Pith is reading between the lines

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

  • EMG's reliance on paired expert trajectories is the clearest constraint; a natural extension would generate pseudo-expert trajectories by having a strong model or planner solve each training task once, or by using the successful branch of a tree search, removing the paired-data requirement while keeping the graph-edit machinery intact.
  • The graph edit path could be used as a supervision signal for fine-tuning rather than just prompting: labeling the erroneous action segment with its edit operation gives a direct training target for policy learning or for learning to predict corrections.
  • The memory graph's cross-task edges suggest a testable extension: if the query embedding space is replaced with task embeddings learned from the graph structures themselves (e.g., graph kernels or graph neural networks), cross-task transfer might improve on tasks where semantic similarity is misleading.
  • The paper compares against iterative reflection; an untested but plausible consequence is that EMG and test-time reflection are complementary, because for genuinely novel failures not covered by training tasks, retrieved memory may be stale, and a fallback loop could use EMG's edit-path insights to seed a single reflection step.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes Experience Memory Graph (EMG), a memory framework for LLM agents. For each training task, EMG assumes a failed exploration trajectory and a successful expert trajectory, converts both into directed action-decision graphs, and uses Fused Gromov-Wasserstein graph matching to extract a common subgraph and a graph edit path. These are summarized into intra-task node insights; additionally, similar tasks are linked with cross-task edge insights. At test time, the agent retrieves top-K node/edge insights and executes once, without iterative self-reflection. Experiments on ALFWorld and ScienceWorld compare EMG with ReAct, Reflexion, ExpeL, CDMem, and MemP, reporting higher success rate and average reward on seen and unseen splits, plus ablations and a time-cost comparison.

Significance. If the empirical claims hold, EMG is a meaningful departure from the reflect-replay paradigm: it moves error correction to an offline, deterministic graph-computation stage and offers a concrete way to inject successful-task knowledge into small models. The graph construction and edit-path algorithms are stated explicitly (Algorithms 1–3), and the robustness discussion in Appendix C is honest about solver suboptimality. However, the headline comparison is currently under-supported because the method receives oracle expert trajectories for every training task while the baselines do not, and because the reported gains are not accompanied by error bars or multiple seeds. The central idea is promising, but the paper's main empirical claim needs additional controls.

major comments (4)
  1. [§4.1.1, Table 2, §5.3] The comparison conflates graph-edit-path extraction with having oracle demonstrations. EMG assumes a successful expert trajectory τ*_i for every training task (Section 4.1.1), and every node insight, common subgraph, and edit path is computed against that gold trajectory. The baselines (Reflexion, ExpeL, CDMem, MemP) are not given access to successful expert trajectories; they reflect over their own rollouts. Table 2 therefore compares 'EMG with gold demonstrations + graph editing' against 'baselines without gold demonstrations'. The w/o-NE ablation is simply ReAct (no memory), and w/o-E still uses node memories built from expert trajectories, so neither isolates the value of the graph edit path from the value of having a target expert trajectory. To support the central claim, the authors should add (a) a baseline that retrieves the nearest expert trajectory (or uses it as a few-shot exe
  2. [Eq. (3), Appendix C, contribution statement] The abstract, introduction, and contribution bullet describe the method as computing the 'shortest graph edit path', and the text emphasizes 'algorithmic exactness'. However, Eq. (3) relies on the FGWAlign solver, which Appendix C concedes may produce suboptimal node matchings and thus 'a longer path (i.e., more edit operations) than the shortest possible one'. Algorithm 3 always yields a legal edit path, but not necessarily the shortest one. This does not invalidate the method, but it is a mismatch between the claimed determinism/exactness and what is actually computed. The authors should replace 'shortest' with 'a legal edit path' throughout, or alternatively report the actual edit-distance quality (e.g., recall against a known optimal path) for the trajectory graphs used in the experiments.
  3. [Table 2] No error bars, confidence intervals, or multiple-seed results are reported. Several of the key gains are within binomial noise. For example, on ScienceWorld with Qwen3-4B, Seen SR for EMG is 17.53% versus 14.95% for Reflexion and 13.40% for ExpeL; with N=194 in the seen split, the standard error is about 2.7–2.9 percentage points, so those differences are less than one standard error. Similarly, Unseen SR differences (16.11% vs. 11.85%) are also borderline. The paper should report means over at least three seeds and/or Wilson confidence intervals for SR and AR in Table 2 and Table 3, and state the number of independent runs.
  4. [§5.4] The one-shot versus iterative comparison is fair only if the baselines are allowed the same information advantage. Section 5.4 allows Reflexion, ExpeL, and CDMem five test-time iterations while EMG executes once, but EMG's single execution is backed by offline access to gold expert trajectories for every training task. A more informative comparison would also give the iterative baselines access to the same successful trajectory pool (e.g., as retrievable exemplars or as additional few-shot content) while retaining their iterative loop. As written, the 'one-shot vs. iterative' claim is entangled with the oracle-expert asymmetry noted above.
minor comments (5)
  1. [Table 2] The metric order under the column spans is inconsistent: ALFWorld uses S-SR, U-SR, S-AR, while ScienceWorld uses S-SR, U-AR, U-SR. This makes the table harder to read and should be unified.
  2. [§5.6] The sensitivity analysis varies only the number of retrieved nodes top-K. Other free parameters — the edge-connectivity threshold τ, top-k edges per node, and the node relabeling costs C_ij in Eq. (3) — are not studied. The conclusion that EMG is 'not sensitive to the exact choice of k' is fine, but it should not be generalized to the other hyperparameters.
  3. [Eq. (4)] The symbol P is overloaded: P denotes both the hard matching matrix and the set of permutation matrices. Use a different symbol (e.g., M or 𝒫) for one of them.
  4. [Figure 3 and Figure 4] The figures report point estimates without error bars. Also, the caption of Figure 3 uses 'Memp' while the text uses 'MemP'; please standardize.
  5. [§4.3] The retrieval formula R(q_test) collects insights from top-K nodes and their top-T neighbors, but the values of K and T used in the experiments are not clearly stated in the main text; they appear only in the sensitivity analysis. Please report the exact retrieval settings used for Table 2.

Circularity Check

0 steps flagged

No significant circularity: EMG's correction paths are definitional graph edits, and its empirical claim rests on held-out test tasks.

full rationale

The paper's derivation chain is not circular. EMG defines error correction as the graph edit path that transforms a failed exploration graph into a paired expert graph (Section 4.2.1, Algorithms 2–3), so the correction path is by construction a function of both inputs; the paper nowhere claims to predict the expert trajectory without using it. The test-time claim is evaluated on held-out ALFWorld/ScienceWorld splits, where retrieval (Eq. 7) selects training-task insights by embedding similarity to q_test, not from the test task's gold trajectory. The expert-trajectory assumption in Section 4.1.1 is an explicit input and is acknowledged as a limitation with expert-free settings deferred to future work; this is an oracle-access condition on the method, not a circular reduction. There are no load-bearing self-citations: external works (ETO for expert data, FGWAlign for graph matching) are used as tools/data sources. The seen-split results are partly memorization by design, and the baseline comparison may be confounded by expert-trajectory availability, but those are benchmark-design concerns rather than instances where a stated result is equivalent to its inputs by construction.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 0 invented entities

EMG's construction depends on gold expert trajectories for every training task, unstated hyperparameters (tau, K, T, embedding model, relabel cost), and an LLM summarization stage. These are not free parameters fitted to test data, but they are unexamined choices on which the empirical claim rests.

free parameters (6)
  • Top-K retrieved nodes = not reported (varied 2-9 in sensitivity)
    Section 4.3 and 5.6; retrieval count K affects how many node insights enter the prompt; main experiments do not state K.
  • Top-T neighbor insights = not reported
    Section 4.3; number of cross-task edge insights per node selected at test time.
  • Edge-connectivity threshold tau and top-k edges = not reported
    Section 4.2.2; edges are added only if cosine similarity >= tau and among top-k most similar; values omitted.
  • Node relabeling cost C_ij = unspecified
    Equation 3; required by the FGW objective but its construction (0/1 or semantic distance) is not specified.
  • Query embedding model = not named
    Sections 4.2.2 and 4.3; all task similarities depend on the embedding model, which the paper does not identify.
  • LLM summarization prompts = not included
    Equations 5-6; intra- and cross-task insights are generated by prompting an LLM, but neither prompt is shown; this is a major uncontrolled component.
axioms (6)
  • domain assumption A successful expert trajectory is available for every training task.
    Section 4.1.1 assumes access to tau*_i via ETO; the method collapses without it.
  • domain assumption Embedding similarity between task queries implies transferable expert workflows.
    Retrieval assumes cosine similarity identifies tasks whose expert action graphs share correct sub-structures; not otherwise validated.
  • domain assumption Normalized action tuples and observation labels preserve decision semantics.
    Algorithm 1 uses exact label equality for matching; paraphrases or synonyms are treated as different, which may produce spurious edit operations.
  • standard math FGW optimization and linear assignment produce valid node matchings.
    Equations 3-4 and reference [20]; standard optimal transport machinery, with the caveat that the solver is approximate (Appendix C).
  • standard math Algorithm 3 always generates a legal edit path.
    Appendix C argues that for any node matching, the returned insertions/deletions/relabelings transform the failed graph into the expert graph.
  • domain assumption LLM summarization preserves the corrective content of graph edit paths and common subgraphs.
    Equations 5-6; final insights are natural-language summaries whose fidelity to the structural corrections is not measured.

pith-pipeline@v1.3.0-alltime-deepseek · 17210 in / 15564 out tokens · 149488 ms · 2026-08-02T03:25:00.910992+00:00 · methodology

0 comments
read the original abstract

Large Language Model (LLM) agents have shown remarkable capabilities in autonomous decision-making by generating sequential trajectories of states, actions, and observations. However, in complex, long-horizon tasks, these agents frequently suffer from compounding errors and struggle to recover from failures. Existing self-correction mechanisms rely on prompt-based reflection, which is inherently brittle, incurs heavy time and API costs due to iterative trial-and-error loops, and produces task-specific memory that may be hard to generalize to new scenarios. To address this, we propose Experience Memory Graph (EMG), a framework that reformulates agent failure recovery as a graph matching problem. At training time, we convert both failed exploration trajectories and successful expert trajectories into directed action decision graphs. By matching these graphs, we extract common subgraphs (successful workflows) and graph edit paths that explicitly indicate how to correct failures (e.g., which actions to add, delete, or relabel under a given observation), and store them in a memory graph with intra-task nodes and cross-task edges. At test time, EMG retrieves relevant insights and guides the agent in a single, loop-free execution. Experiments on ALFWorld and ScienceWorld show that EMG consistently outperforms state-of-the-art reflection baselines in success rate and average reward, while requiring no test-time trial-and-error.

Figures

Figures reproduced from arXiv: 2607.13884 by Fengrui Liu, Kai Zheng, Wenjun Wang, Yuchen Fang, Zibo Liang.

Figure 1
Figure 1. Figure 1: Difference of error correction based on (a) reflect [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overall architecture of EMG (Experience Memory Graph). In the offline construction phase, we collect exploration [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 5
Figure 5. Figure 5: Sensitivity analysis of EMG. We report the success [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 4
Figure 4. Figure 4: Time cost of EMG and iterative self-reflection base [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: A concrete example for constructing action decision [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

43 extracted references · 8 linked inside Pith

  1. [1]

    Sorokin, Dmitry Evseev, Andrey Kravchenko, Mikhail Burtsev, and Evgeny Burnaev

    Petr Anokhin, Nikita Semenov, Artyom Y. Sorokin, Dmitry Evseev, Andrey Kravchenko, Mikhail Burtsev, and Evgeny Burnaev. 2025. AriGraph: Learn- ing Knowledge Graph World Models with Episodic Memory for LLM Agents. InProceedings of the Thirty-Fourth International Joint Conference on Artificial Intelligence, IJCAI 2025, Montreal, Canada, August 16-22, 2025. ...

    doi:10.24963/ijcai.2025/2 2025

  2. [2]

    Qihao Cheng, Da Yan, Tianhao Wu, Zhongyi Huang, and Qin Zhang. 2025. Computing approximate graph edit distance via optimal transport.Proceedings of the ACM on Management of Data3, 1 (2025), 1–26

    2025

  3. [3]

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

  4. [4]

    Runnan Fang, Yuan Liang, Xiaobin Wang, Jialong Wu, Shuofei Qiao, Pengjun Xie, Fei Huang, Huajun Chen, and Ningyu Zhang. 2025. Memp: Exploring agent procedural memory.arXiv preprint arXiv:2508.06433(2025)

    Pith/arXiv arXiv 2025

  5. [5]

    Pengyu Gao, Jinming Zhao, Xinyue Chen, and Long Yilin. 2025. An efficient context-dependent memory framework for llm-centric agents. InProceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 3: Industry Track). 1055–1069

    2025

  6. [6]

    Rui Ge, Yichao Fu, Yuyang Qian, Junda Su, Yiming Zhao, Peng Zhao, and Hao Zhang. 2026. Internalizing Agency from Reflective Experience.arXiv preprint arXiv:2603.16843(2026)

    arXiv 2026

  7. [7]

    Yuyang Hu, Shichun Liu, Yanwei Yue, Guibin Zhang, Boyang Liu, Fangyi Zhu, Jiahang Lin, Honglin Guo, Shihan Dou, Zhiheng Xi, et al. 2025. Memory in the age of ai agents.arXiv preprint arXiv:2512.13564(2025)

    Pith/arXiv arXiv 2025

  8. [8]

    Tenghao Huang, Kinjal Basu, Ibrahim Abdelaziz, Pavan Kapanipathi, Jonathan May, and Muhao Chen. 2025. R2d2: Remembering, replaying and dynamic decision making with a reflective agentic memory. InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 30318–30330

    2025

  9. [9]

    Namyoung Kim, Kai Tzu-iunn Ong, Yeonjun Hwang, Minseok Kang, Iiseo Jihn, Gayoung Kim, Minju Kim, and Jinyoung Yeo. 2025. PRINCIPLES: Synthetic Strategy Memory for Proactive Dialogue Agents. InThe 2025 Conference on Empirical Methods in Natural Language Processing (EMNLP 2025)

    2025

  10. [10]

    Canasai Kruengkrai and Koichiro Yoshino. 2025. Teaching Text Agents to Learn Sequential Decision Making from Failure. InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 31619–31635

    2025

  11. [11]

    Xiaoxi Li, Wenxiang Jiao, Jiarui Jin, Guanting Dong, Jiajie Jin, Yinuo Wang, Hao Wang, Yutao Zhu, Ji-Rong Wen, Yuan Lu, et al . 2026. Deepagent: A general reasoning agent with scalable toolsets. InProceedings of the ACM Web Conference

    2026

  12. [12]

    Xuechen Liang, Meiling Tao, Yinghui Xia, Jianhui Wang, Kun Li, Yijin Wang, Yangfan He, Jingsong Yang, Tianyu Shi, Yuantao Wang, et al. 2025. Sage: Self- evolving agents with reflective and memory-augmented abilities.Neurocomputing 647 (2025), 130470

    2025

  13. [13]

    Yixin Liu, Guibin Zhang, Kun Wang, Shiyuan Li, Shirui Pan, and Bo An. 2026. Graph-augmented large language model agents: Current progress and future prospects.IEEE Intelligent Systems41, 2 (2026), 45–55

    2026

  14. [14]

    Qirui Mi, Zhijian Ma, Mengyue Yang, Haoxuan Li, Yisen Wang, Haifeng Zhang, and Jun Wang. 2026. Skill-Pro: Learning Reusable Skills from Experience via Non-Parametric PPO for LLM Agents. InProceedings of the 43rd International Conference on Machine Learning (ICML 2026). https://arxiv.org/abs/2602.01869 Spotlight

    Pith/arXiv arXiv 2026

  15. [15]

    Siru Ouyang, Jun Yan, I Hsu, Yanfei Chen, Ke Jiang, Zifeng Wang, Rujun Han, Long T Le, Samira Daruki, Xiangru Tang, et al. 2025. Reasoningbank: Scaling agent self-evolving with reasoning memory.arXiv preprint arXiv:2509.25140 (2025)

    Pith/arXiv arXiv 2025

  16. [16]

    Revanth Gangi Reddy, Sagnik Mukherjee, Jeonghwan Kim, Zhenhailong Wang, Dilek Hakkani-Tur, and Heng Ji. 2025. Infogent: An agent-based framework for web information aggregation. InFindings of the Association for Computational Linguistics: NAACL 2025. 5745–5758

    2025

  17. [17]

    Jie-Jing Shao, Haiyan Yin, Yueming Lyu, Xingrui Yu, Lan-Zhe Guo, Ivor Tsang, James Kwok, and Yu-Feng Li. 2026. Lifting Traces to Logic: Programmatic Skill Induction with Neuro-Symbolic Learning for Long-Horizon Agentic Tasks.arXiv preprint arXiv:2605.01293(2026)

    Pith/arXiv arXiv 2026

  18. [18]

    Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. 2023. Reflexion: Language agents with verbal reinforcement learning. Advances in neural information processing systems36 (2023), 8634–8652

    2023

  19. [19]

    Yifan Song, Da Yin, Xiang Yue, Jie Huang, Sujian Li, and Bill Yuchen Lin. 2024. Trial and error: Exploration-based trajectory optimization of LLM agents. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 7584–7600

    2024

  20. [20]

    Jianheng Tang, Xi Zhao, Lemin Kong, Xiaofang Zhou, and Jia Li. 2025. Fused Gromov-Wasserstein Alignment for Graph Edit Distance Computation and Be- yond.Proceedings of the VLDB Endowment18, 10 (2025), 3641–3654

    2025

  21. [21]

    Hanlin Wang, Jian Wang, Chak Tou Leong, and Wenjie Li. 2025. Steca: Step-level trajectory calibration for llm agent learning. InFindings of the Association for Computational Linguistics: ACL 2025. 11597–11614

    2025

  22. [22]

    Zora Zhiruo Wang, Jiayuan Mao, Daniel Fried, and Graham Neubig. 2025. Agent Workflow Memory. InInternational Conference on Machine Learning. PMLR, 63897–63911

    2025

  23. [23]

    Zhepei Wei, Wenlin Yao, Yao Liu, Weizhi Zhang, Qin Lu, Liang Qiu, Changlong Yu, Puyang Xu, Chao Zhang, Bing Yin, et al. 2025. Webagent-r1: Training web agents via end-to-end multi-turn reinforcement learning. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 7920–7939

    2025

  24. [24]

    Yaxiong Wu, Yongyue Zhang, Sheng Liang, and Yong Liu. 2025. Sgmem: Sentence graph memory for long-term conversational agents.arXiv preprint arXiv:2509.21212(2025)

    arXiv 2025

  25. [25]

    Zhaofen Wu, Hanrong Zhang, Fulin Lin, Wujiang Xu, Xinran Xu, Yankai Chen, Henry Peng Zou, Shaowen Chen, Weizhi Zhang, Xue Liu, et al. 2026. Gam: Hierar- chical graph-based agentic memory for llm agents.arXiv preprint arXiv:2604.12285 (2026)

    Pith/arXiv arXiv 2026

  26. [26]

    Siyu Xia, Zekun Xu, Jiajun Chai, Wentian Fan, Yan Song, Xiaohan Wang, Guo- jun Yin, Wei Lin, Haifeng Zhang, and Jun Wang. 2025. From experience to strategy: Empowering llm agents with trainable graph memory.arXiv preprint arXiv:2511.07800(2025)

    arXiv 2025

  27. [27]

    Tianle Xia, Lingxiang Hu, Yiding Sun, Ming Xu, Lan Xu, Siying Wang, Wei Xu, and Jie Jiang. 2026. GraSP: Graph-Structured Skill Compositions for LLM Agents

    2026

  28. [28]

    Wujiang Xu, Zujie Liang, Kai Mei, Hang Gao, Juntao Tan, and Yongfeng Zhang

  29. [29]

    Wei Yang, Jinwei Xiao, Hongming Zhang, Qingyang Zhang, Yanna Wang, and Bo Xu. 2025. Coarse-to-fine grounded memory for llm agent planning.arXiv preprint arXiv:2508.15305(2025)

    arXiv 2025

  30. [30]

    A-mem: Agentic memory for llm agents.Advances in Neural Information Processing Systems38 (2026), 17577–17604

    2026

  31. [31]

    Tianjun Yao, Yongqiang Chen, Yujia Zheng, Pan Li, Zhiqiang Shen, and Kun Zhang. 2026. ParamMem: Augmenting Language Agents with Parametric Reflec- tive Memory.arXiv preprint arXiv:2602.23320(2026)

    arXiv 2026

  32. [32]

    Narasimhan, and Yuan Cao

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik R. Narasimhan, and Yuan Cao. 2023. ReAct: Synergizing Reasoning and Acting in Language Models. InThe Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net. https://openreview. net/forum?id=WE_vluYUL-X

    2023

  33. [33]

    Jiawei Yu, Yixiang Fang, Xilin Liu, and Yuchi Ma. 2026. H-Mem: A Novel Memory Mechanism for Evolving and Retrieving Agent Memory via a Hybrid Structure. arXiv preprint arXiv:2605.15701(2026)

    Pith/arXiv arXiv 2026

  34. [34]

    Weiran Yao, Shelby Heinecke, Juan Carlos Niebles, Zhiwei Liu, Yihao Feng, Le Xue, Rithesh Ramapura Narasimha Murthy, Zeyuan Chen, Jianguo Zhang, Devansh Arpit, et al. 2024. Retroformer: Retrospective large language agents with policy gradient optimization. InInternational Conference on Learning Representations, Vol. 2024. 10091–10111

    2024

  35. [35]

    Guibin Zhang, Muxin Fu, Kun Wang, Frank Wan, Miao Yu, and Shuicheng Yan

  36. [36]

    Dengjia Zhang, Xiaoou Liu, Lu Cheng, Yaqing Wang, Kenton Murray, and Hua Wei. 2026. SELAUR: Self Evolving LLM Agent via Uncertainty-aware Rewards. arXiv preprint arXiv:2602.21158(2026)

    arXiv 2026

  37. [37]

    Andrew Zhao, Daniel Huang, Quentin Xu, Matthieu Lin, Yong-Jin Liu, and Gao Huang. 2024. Expel: Llm agents are experiential learners. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 38. 19632–19642

    2024

  38. [38]

    G-memory: Tracing hierarchical memory for multi-agent systems.Advances in Neural Information Processing Systems38 (2026), 12988–13018

    2026

  39. [39]

    Guibin Zhang, Junhao Wang, Junjie Chen, Wangchunshu Zhou, Kun Wang, and Shuicheng Yan. 2025. AgenTracer: Who Is Inducing Failure in the LLM Agentic Systems?arXiv preprint arXiv:2509.03312(2025)

    Pith/arXiv arXiv 2025

  40. [40]

    nothing happened

    Kunlun Zhu, Zijia Liu, Bingxuan Li, Muxin Tian, Yingxuan Yang, Jiaxun Zhang, Pengrui Han, Qipeng Xie, Fuyang Cui, Weijia Zhang, et al . 2025. Where llm agents fail and how they can learn from failures.arXiv preprint arXiv:2509.25370 (2025). KDD ’27, August 1–5, 2027, San Jose, United States. Wenjun Wang et al. Figure 6: A concrete example for constructing...

    arXiv 2025

  41. [41]

    Longtao Zheng, Rundong Wang, Xinrun Wang, and Bo An. 2024. Synapse: Trajectory-as-exemplar prompting with memory for computer control. InInter- national Conference on Learning Representations, Vol. 2024. 19036–19066

    2024

  42. [42]

    Xu Zheng, Zhuomin Chen, Chaohao Lin, Hua Wei, Haifeng Chen, Wei Cheng, and Dongsheng Luo. 2026. Trajectory Graph Copilot: Pre-Action Error Diagnosis in LLM Agents. https://openreview.net/forum?id=ighxnB6nJF

    2026

  43. [2026]

    InThe Fourteenth International Conference on Learning Representations

    LightMem: Lightweight and Efficient Memory-Augmented Generation. InThe Fourteenth International Conference on Learning Representations. https: //openreview.net/forum?id=dyJ0GWpjJB