Pith. sign in

REVIEW 5 major objections 5 minor 3 cited by

ActMem claims that an LLM agent's memory should be a causal graph, not a text archive, and that counterfactual retrieval over this graph outperforms retrieval-heavy baselines on memory-dependent reasoning tasks.

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

2026-08-03 04:43 UTC pith:6GLER4ON

load-bearing objection ActMem is a serious, well-motivated agent-memory framework whose headline gains on its own benchmark are plausible but not yet airtight; the evaluation loop needs to be stress-tested before believing the margins. the 5 major comments →

arxiv 2603.00026 v2 pith:6GLER4ON submitted 2026-02-04 cs.CL cs.AIcs.IR

ActMem: Bridging the Gap Between Memory Retrieval and Reasoning in LLM Agents

Xiaohui Zhang , Zequn Sun , Chengyuan Yang , Yaqin Jin , Yazhong Zhang , Wei Hu This is my paper
classification cs.CL cs.AIcs.IR
keywords memory-augmented agentscausal reasoningcounterfactual reasoningknowledge graphLLM agentslong-term memoryretrieval-augmented generationcommonsense completion
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.

Current memory frameworks treat past dialogue as passive text to be retrieved by similarity; ActMem instead compresses it into atomic facts and builds a memory knowledge graph with both semantic and causal edges. At query time it first asks what negative consequence the user's intention might have, then uses that counterfactual commonsense statement to retrieve the relevant past. The paper claims this closes the gap between memory retrieval and reasoning: it outperforms state-of-the-art baselines by 12.55 percentage points in QA accuracy on its new ActMemEval benchmark with a strong LLM backend, and by about 7 points on a standard long-term memory benchmark. The core insight is that retrieval accuracy alone is not decision utility; reasoning with memory requires detecting implicit constraints and conflicts.

Core claim

The central discovery is that high retrieval accuracy does not imply good decisions. ActMem transforms unstructured dialogue into a knowledge graph where facts are linked by semantic edges (topic similarity) and causal edges (LLM-proposed causes, filtered by a pointwise mutual information score). Retrieval is a two-step loop: an initial vector search, then a counterfactual question ('if the user does this, what bad thing could happen?') whose answer becomes an enhanced query to pull in the hidden evidence. This lets the agent infer, for example, that buying Sago Palms conflicts with owning a chewing puppy, and respond with a safety warning instead of a store recommendation. Ablations confirm

What carries the argument

The memory knowledge graph (memory KG). Nodes are atomic facts extracted from dialogue via an LLM; edges are either semantic (cosine similarity above a threshold) or causal (an LLM proposes candidate edges, and a pointwise mutual information score based on conditional language-model loss keeps only those where the cause reduces uncertainty about the effect). The retrieval-and-reasoning loop uses the graph in two passes: retrieve by similarity, generate a counterfactual commonsense statement, then retrieve again against that statement. The interplay of the two edge types and the counterfactual step is what carries the argument.

Load-bearing premise

The framework assumes that the LLM-proposed causal edges, after the PMI filter, are genuine cause-effect relations; since PMI measures statistical association rather than causation, a spurious correlation could be accepted and mislead the counterfactual retrieval.

What would settle it

Construct a dialogue where two events co-occur frequently but have no causal link (e.g., the user always talks about the puppy right before mentioning the weather). If ActMem treats the weather as caused by the puppy and uses that edge to answer a question, the framework will produce a wrong or overconfident response. More directly, run ActMem on a benchmark where all conflict scenarios are hand-authored by humans rather than LLM-synthesized; if the accuracy gap over retrieval baselines vanishes, the edge filters were encoding the benchmark's own inductive bias.

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

If this is right

  • If ActMem is right, memory management for agents should be evaluated by decision quality or conflict detection, not by recall accuracy alone.
  • Agents built this way could proactively warn users about dangerous or inconsistent actions (e.g., buying a toxic plant when a puppy is at home).
  • The framework's QA gains transfer to an existing long-term memory benchmark, suggesting the causal-graph structure helps factual retrieval too, not just reasoning.
  • The new ActMemEval benchmark provides a template for constructing logic-driven memory tasks with synthetic causal topologies and noise injection.

Where Pith is reading between the lines

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

  • The counterfactual-retrieval pattern might generalize to other agent decisions—tool selection, planning, or refusal—where hidden constraints live in past interactions; that is not tested in the paper.
  • The PMI filter only measures association, so a same-topic but non-causal pair could pass; a stricter causal test (e.g., intervention-based or human-annotated causal edges) would either confirm or narrow the claimed advantage.
  • Because ActMemEval's ground truths are generated and validated by LLMs, the benchmark may share blind spots with the models being evaluated; a fully human-authored conflict set would be a stronger falsifier.
  • A testable extension: use ActMem's graph to explain decisions by returning the causal chain, which could make agent behavior more inspectable.

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

5 major / 5 minor

Summary. The paper proposes ActMem, a memory-management framework for LLM agents that converts dialogue history into a knowledge graph of memory facts with semantic and causal edges and performs counterfactual reasoning to retrieve implicit constraints. The authors also introduce ActMemEval, a synthetic benchmark of 246 logic-driven scenarios, and report experiments on ActMemEval and LongMemEval showing QA accuracy gains over several baselines, including a 12.55 pp margin over LightMem with DeepSeek-V3. The core claim is that ActMem significantly outperforms state-of-the-art memory frameworks on reasoning-over-memory tasks.

Significance. The paper targets a genuine limitation of current memory systems and proposes a concrete, mechanism-based framework plus a new benchmark. If the empirical claims hold, the contribution would be relevant to LLM-agent memory research. The ablation study is informative and the dataset construction pipeline is a useful step beyond purely retrieval-focused benchmarks. However, the current evidence is weakened by a partly self-referential evaluation loop, lack of statistical validation, and an unvalidated PMI-based causality criterion. The central comparative claim is plausible but not yet convincingly supported.

major comments (5)
  1. [§5.2, Table 1] The QA accuracy is computed by LLM-as-a-Judge using DeepSeek-V3, which is also the backbone for the ActMem system in the same table. Because ActMem's responses explicitly present structured causal/counterfactual reasoning, they may stylistically resemble the LLM-generated gold answers more than baseline outputs do, inflating the reported margin. No judge prompt, rubric, or human agreement is reported. Please add a human-annotated sample (with agreement) or use a judge from a different model family, and compare against an equally reasoning-style baseline.
  2. [§4.1, Steps 1–5] ActMemEval is generated by an LLM, filtered by an LLM validator, and only manually checked for naturalness and inferability. The manual verification is described without annotator counts, instructions, or inter-annotator agreement. If the benchmark encodes the same linguistic and commonsense patterns that ActMem's LLM components rely on, the advantage over RAG may be partially due to this alignment. Please provide details of the verification and report human performance on a sample or a negative-control experiment.
  3. [§5.4, Table 1] The abstract and §5.4 claim that ActMem 'significantly outperforms' baselines, but no confidence intervals, significance tests, or multiple-run variation are reported. All results are single numbers. At minimum, run each configuration multiple times with different seeds and apply a paired significance test to support the 'significant' wording.
  4. [§3.3, Eqs. (4)–(5)] The PMI filter measures the reduction in cross-entropy of a pretrained language model when conditioning on a candidate antecedent. This is a measure of textual association, not causal intervention, yet the resulting edges are called 'causal' and the framework's reasoning mechanism is built on them. The paper supplies no validation that this criterion corresponds to human judgments of causality. Please either validate the PMI criterion against human-annotated causal relations or weaken the causal claims.
  5. [§5.7, Table 2] The semantic-edge similarity threshold τ is chosen based on QA accuracy on the same ActMemEval test set (0.3 gives 76.52%). This is a form of test-set tuning. The reported sensitivity analysis does not justify the chosen threshold without a validation split or a clear pre-registered selection criterion.
minor comments (5)
  1. [Table 1] The naming 'A-mem' and 'A-Mem' is inconsistent between the table and the text. Also, MemoryOS lacks retrieval-accuracy entries; explain in the table caption that this is because memories could not be extracted, not just in a footnote.
  2. [§4.2, Figure 4] The figure shows a distribution of semantic similarity, but the embedding model used to compute similarities is not specified. Please add this information in the caption or text.
  3. [§5.2] The retrieval-accuracy metric is said to follow HaluMem [Chen et al., 2025], but how the ground-truth evidence facts are defined and obtained for ActMemEval is not explicitly stated. Please clarify the annotation of evidence facts.
  4. [§5.8, Figure 7] The bar chart has no error bars or significance markers, so the 6.96 pp gain over LightMem lacks uncertainty quantification. Please add error bars if any runs are repeated.
  5. [Throughout] The phrase 'state-of-the-art baselines' is used for methods that are not always evaluated in comparable settings (e.g., A-Mem and MemoryOS had missing results). Please temper the phrasing or provide complete baseline results.

Circularity Check

2 steps flagged

ActMemEval's headline QA margin is partly self-referential: the semantic threshold is tuned on the same benchmark, and the LLM-generated benchmark is judged by the same model family as the ActMem backbone; LongMemEval provides independent but weaker support.

specific steps
  1. fitted input called prediction [§5.3 Implementation Details; §5.7 Analysis of Semantic Similarity Thresholds (Table 2 vs Table 1)]
    "The threshold used for semantic-edge filtering is 0.3. ... A lower threshold (e.g., 0.3) yields the optimal QA accuracy of 76.52%."

    The main-result QA accuracy 76.52% (Table 1) is not produced by a fixed, pre-specified configuration: it is the maximum of a threshold sweep over the same ActMemEval set. The final semantic-edge threshold was chosen because it gave this exact accuracy, so the headline ActMem-vs-LightMem margin on ActMemEval is partly forced by test-set fitting rather than by the memory mechanism.

  2. other [§4.1 Steps 1–2; §5.2 Evaluation Metrics; §5.3 Implementation Details]
    "We employ an advanced LLM as the reasoning topology synthesizer. ... we utilize an LLM-as-a-Judge method (using DeepSeek-V3) to score the correctness of the response against the ground truth. ... We employ two LLMs as backbones for reasoning and generation: DeepSeek-V3, representing high-performance open-source models."

    ActMemEval's gold answers are synthesized by an LLM and validated by an LLM, and correctness is judged by DeepSeek-V3, which is also the ActMem backbone. Because ActMem deliberately produces structured causal/counterfactual reasoning, a DeepSeek judge may systematically prefer reasoning-style text that resembles the LLM-written reference answers, inflating the reported 12.55pp margin over LightMem. This is an evaluation-loop self-reference rather than an algebraic identity, but it is load-bearing for the headline claim.

full rationale

The paper's central mechanism is not a formal derivation, so strict self-definitional circularity is limited. The clearest concrete leak is the semantic-edge threshold: §5.3 fixes it to 0.3, and §5.7 shows 0.3 was selected because it maximizes QA accuracy on the same ActMemEval set; the reported 76.52% is therefore a selected optimum, not an independent prediction. In addition, the benchmark and judge are internal to the LLM family being tested: ground truth comes from an LLM topology synthesizer and validator, and the judge is DeepSeek-V3, the same family as the ActMem backbone, so the evaluation loop contains the method it evaluates. However, I do not score this as fully circular because (a) LongMemEval provides an external-benchmark result where ActMem still beats LightMem by 6.96pp, and (b) ablations isolate the causal/semantic/reasoning modules. Those external and ablation results give the central claim independent content, so the circularity is partial and concentrated in the self-built benchmark and the threshold fit.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 0 invented entities

No new physical entities are postulated. The central claim depends on several hand-set thresholds and unverified assumptions about LLM causal extraction and judge reliability.

free parameters (5)
  • clustering distance threshold = 0.2
    Used in single-pass incremental clustering (Section 3.2); chosen without sensitivity analysis.
  • semantic-edge similarity threshold = 0.3
    Filters semantic edges; Table 2 shows this value gives the highest QA accuracy, so it is effectively tuned on ActMemEval.
  • PMI validation threshold = 0.8
    Retains causal edges with S_PMI > tau (Eq. 5); described as 'empirically set' in Section 5.3, with no sensitivity study.
  • initial retrieval count = 20
    Top-k facts for the first retrieval step (Section 5.3); chosen by hand.
  • counterfactual retrieval count = 10
    Top-k for the refinement step (Section 5.3); chosen by hand.
axioms (6)
  • domain assumption The LLM information extractor produces accurate, atomic memory facts from raw dialogue.
    Section 3.1 assumes fext reliably compresses history without losing key entities/events needed for later reasoning.
  • domain assumption Embedding cosine similarity captures topical and semantic relatedness between facts.
    Sections 3.2–3.3 use Qwen3-Embedding cosine similarity to cluster facts and add semantic edges.
  • ad hoc to paper PMI reduction in GPT2-Large cross-entropy indicates causal dependency.
    Eq. 4 defines S_PMI as loss reduction when prefixing fi; no independent evidence distinguishes causation from mere correlation.
  • ad hoc to paper The counterfactual LLM prompt produces useful commonsense knowledge kcs.
    Section 3.4 Step 2 relies on the prompt 'If the user does q, what negative consequences might occur' to bridge query and memory; correctness is not verified.
  • domain assumption DeepSeek-V3 as LLM-as-a-Judge correctly scores open-ended answers.
    Section 5.2 uses DeepSeek-V3 as judge; no human-agreement study is reported, and the same model family is used as a backbone.
  • domain assumption The synthetic ActMemEval topologies have valid, unambiguous logical ground truth.
    Section 4.1 Step 2 filters topologies with an LLM validator; manual verification covers the final 246 samples but does not independently validate the LLM-generated logic.

pith-pipeline@v1.3.0-alltime-deepseek · 11979 in / 10110 out tokens · 94084 ms · 2026-08-03T04:43:20.061557+00:00 · methodology

0 comments
read the original abstract

Memory management is essential for LLM agents in long-term interactions. Current memory frameworks typically treat agents as passive ``recorders'' and retrieve information without understanding its deeper implications. They may fail in scenarios requiring reasoning and complex decision-making. To bridge this critical gap, we propose a novel actionable memory framework called ActMem that integrates memory retrieval with active causal reasoning. ActMem transforms unstructured dialogue history into a structured causal and semantic graph. By leveraging counterfactual reasoning and commonsense completion, it enables agents to deduce implicit constraints and resolve potential conflicts between past states and current intentions. Furthermore, we introduce a comprehensive dataset ActMemEval to evaluate agent reasoning capabilities in logic-driven scenarios, moving beyond the fact-retrieval focus of existing memory benchmarks. Experiments demonstrate that ActMem significantly outperforms baselines in handling complex, memory-dependent tasks, paving the way for more consistent and reliable intelligent assistants.

Figures

Figures reproduced from arXiv: 2603.00026 by Chengyuan Yang, Wei Hu, Xiaohui Zhang, Yaqin Jin, Yazhong Zhang, Zequn Sun.

Figure 1
Figure 1. Figure 1: A motivating comparison between memory retrieval and [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The overview of our proposed framework. complex reasoning capabilities. Recent initiatives, includ￾ing RULER [Hsieh et al., 2024] and BABILong [Kuratov et al., 2024] include variable tracking and logical chaining, while other datasets like LoCoMo [Maharana et al., 2024] and LongMemEval [Wu et al., 2025] target long-term dialogue consistency and agent memory retrieval. Despite these ad￾vancements, existing … view at source ↗
Figure 3
Figure 3. Figure 3: Pipeline to construct our ActMemEval dataset. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Distribution of semantic similarity scores between the [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Ablation study on counterfactual reasoning, causal edge [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: Performance comparison on LongMemEval. more time than the summarization-based LightMem to con￾struct causal graphs, it remains significantly faster than other graph-based baselines. The second and third subfigures high￾light the cost-effectiveness. For the number of input tokens, ActMem consumes a moderate amount to extracting fine￾grained causal events. Regarding the number of output tokens, ActMem achiev… view at source ↗

discussion (0)

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

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. What Happens Inside Agent Memory? Circuit Analysis from Emergence to Diagnosis

    cs.AI 2026-05 unverdicted novelty 6.0

    Circuit analysis reveals that routing circuits for agent memory emerge at 0.6B parameters while content circuits emerge at 4B, with a shared grounding hub and an unsupervised diagnostic achieving 76.2% accuracy for lo...

  2. What Happens Inside Agent Memory? Circuit Analysis from Emergence to Diagnosis

    cs.AI 2026-05 unverdicted novelty 6.0

    In LLM agents, memory routing circuits emerge at 0.6B scale while content circuits appear only at 4B, and write/read operations recruit a pre-existing late-layer context hub instead of creating a new one, enabling a 7...

  3. ConMem: Contribution-Aware Memory for Long-Horizon Manufacturing Inspection Logs

    cs.AI 2026-07 conditional novelty 5.5

    Role-segmented inspection memories ranked by approximated Shapley contribution yield 76% early-risk QA accuracy with ~88% fewer tokens than naive 8K-context LLMs.

Reference graph

Works this paper leans on

24 extracted references · 13 linked inside Pith · cited by 2 Pith papers

  1. [1]

    HaluMem: Evaluating hallucinations in memory systems of agents.CoRR, abs/2511.03506,

    [Chenet al., 2025 ] Ding Chen, Simin Niu, Kehang Li, Peng Liu, Xiangping Zheng, Bo Tang, Xinchi Li, Feiyu Xiong, and Zhiyu Li. HaluMem: Evaluating hallucinations in memory systems of agents.CoRR, abs/2511.03506,

    arXiv 2025

  2. [3]

    LightMem: Lightweight and efficient memory-augmented generation

    [Fanget al., 2025 ] Jizhan Fang, Xinle Deng, Haoming Xu, Ziyan Jiang, Yuqi Tang, Ziwen Xu, Shumin Deng, Yun- zhi Yao, Mengru Wang, Shuofei Qiao, et al. LightMem: Lightweight and efficient memory-augmented generation. CoRR, abs/2510.18866,

    Pith/arXiv arXiv 2025

  3. [4]

    Retrieval-augmented generation for large language models: A survey.CoRR, abs/2312.10997,

    [Gaoet al., 2023 ] Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, Qianyu Guo, Meng Wang, et al. Retrieval-augmented generation for large language models: A survey.CoRR, abs/2312.10997,

    Pith/arXiv arXiv 2023

  4. [6]

    RULER: What’s the real context size of your long-context language models? InCoLM,

    [Hsiehet al., 2024 ] Cheng-Ping Hsieh, Simeng Sun, Samuel Kriman, Shantanu Acharya, Dima Rekesh, Fei Jia, and Boris Ginsburg. RULER: What’s the real context size of your long-context language models? InCoLM,

    2024

  5. [8]

    Memory OS of AI agent.CoRR, abs/2506.06326,

    [Kanget al., 2025 ] Jiazheng Kang, Mingming Ji, Zhe Zhao, and Ting Bai. Memory OS of AI agent.CoRR, abs/2506.06326,

    Pith/arXiv arXiv 2025

  6. [9]

    RaDA: Retrieval-augmented web agent planning with LLMs

    [Kimet al., 2024 ] Minsoo Kim, Victor Bursztyn, Eunyee Koh, Shunan Guo, and Seung won Hwang. RaDA: Retrieval-augmented web agent planning with LLMs. In Findings of the Association for Computational Linguistics: ACL 2024, pages 13511–13525,

    2024

  7. [10]

    MapAgent: Trajectory-constructed memory- augmented planning for mobile task automation.CoRR, abs/2507.21953,

    [Konget al., 2025 ] Yi Kong, Dianxi Shi, Guoli Yang, Zhang ke-di, Chenlin Huang, Xiaopeng Li, and Songchang Jin. MapAgent: Trajectory-constructed memory- augmented planning for mobile task automation.CoRR, abs/2507.21953,

    Pith/arXiv arXiv 2025

  8. [11]

    In search of needles in a 11M haystack: Recurrent memory finds what LLMs miss.CoRR, abs/2402.10790,

    [Kuratovet al., 2024 ] Yuri Kuratov, Aydar Bulatov, Petr Anokhin, Dmitry Sorokin, Artyom Sorokin, and Mikhail Burtsev. In search of needles in a 11M haystack: Recurrent memory finds what LLMs miss.CoRR, abs/2402.10790,

    Pith/arXiv arXiv 2024

  9. [12]

    MemOS: An operating system for memory-augmented generation (MAG) in large language models.CoRR, abs/2505.22101,

    [Liet al., 2025 ] Zhiyu Li, Shichao Song, Hanyu Wang, Simin Niu, Ding Chen, Jiawei Yang, Chenyang Xi, Huayi Lai, Jihao Zhao, Yezhaohui Wang, et al. MemOS: An operating system for memory-augmented generation (MAG) in large language models.CoRR, abs/2505.22101,

    Pith/arXiv arXiv 2025

  10. [13]

    Evaluating very long-term conversational memory of LLM agents.arxiv,

    [Maharanaet al., 2024 ] Adyasha Maharana, Dong-Ho Lee, Sergey Tulyakov, Mohit Bansal, Francesco Barbieri, and Yuwei Fang. Evaluating very long-term conversational memory of LLM agents.arxiv,

    2024

  11. [14]

    Patil, Kevin Lin, Sarah Wooders, and Joseph E

    [Packeret al., 2023 ] Charles Packer, Vivian Fang, Shishir G. Patil, Kevin Lin, Sarah Wooders, and Joseph E. Gonzalez. MemGPT: Towards LLMs as operating systems.CoRR, abs/2310.08560,

    Pith/arXiv arXiv 2023

  12. [15]

    Zep: A temporal knowledge graph architecture for agent memory

    [Rasmussenet al., 2025 ] Preston Rasmussen, Pavlo Paliy- chuk, Travis Beauvais, Jack Ryan, and Daniel Chalef. Zep: A temporal knowledge graph architecture for agent memory. CoRR, abs/2501.13956,

    Pith/arXiv arXiv 2025

  13. [16]

    From isolated conversations to hierar- chical schemas: Dynamic tree memory representation for LLMs

    [Rezazadehet al., 2025 ] Alireza Rezazadeh, Zichao Li, Wei Wei, and Yujia Bao. From isolated conversations to hierar- chical schemas: Dynamic tree memory representation for LLMs. InICLR,

    2025

  14. [17]

    LongMemEval: Benchmarking chat assistants on long-term interactive memory

    [Wuet al., 2025 ] Di Wu, Hongwei Wang, Wenhao Yu, Yuwei Zhang, Kai-Wei Chang, and Dong Yu. LongMemEval: Benchmarking chat assistants on long-term interactive memory. InICLR,

    2025

  15. [18]

    A-MEM: Agentic memory for LLM agents.CoRR, abs/2502.12110,

    [Xuet al., 2025 ] Wujiang Xu, Zujie Liang, Kai Mei, Hang Gao, Juntao Tan, and Yongfeng Zhang. A-MEM: Agentic memory for LLM agents.CoRR, abs/2502.12110,

    Pith/arXiv arXiv 2025

  16. [19]

    Beyond static summarization: Proactive mem- ory extraction for llm agents.CoRR, abs/2601.04463,

    [Yanget al., 2026 ] Chengyuan Yang, Zequn Sun, Wei Wei, and Wei Hu. Beyond static summarization: Proactive mem- ory extraction for llm agents.CoRR, abs/2601.04463,

    arXiv 2026

  17. [21]

    Qwen3 em- bedding: Advancing text embedding and reranking through foundation models.CoRR, abs/2506.05176,

    [Zhanget al., 2025b ] Yanzhao Zhang, Mingxin Li, Dingkun Long, Xin Zhang, Huan Lin, Baosong Yang, Pengjun Xie, An Yang, Dayiheng Liu, Junyang Lin, et al. Qwen3 em- bedding: Advancing text embedding and reranking through foundation models.CoRR, abs/2506.05176,

    Pith/arXiv arXiv

  18. [22]

    Synapse: Trajectory-as-exemplar prompting with memory for computer control

    [Zhenget al., 2024 ] Longtao Zheng, Rundong Wang, Xin- run Wang, and Bo An. Synapse: Trajectory-as-exemplar prompting with memory for computer control. InICLR. OpenReview.net,

    2024

  19. [23]

    MemoryBank: Enhancing large language models with long-term memory

    [Zhonget al., 2024 ] Wanjun Zhong, Lianghong Guo, Qiqi Gao, He Ye, and Yanlin Wang. MemoryBank: Enhancing large language models with long-term memory. InAAAI, pages 19724–19731,

    2024

  20. [24]

    TRAD: Enhancing LLM agents with step-wise thought retrieval and aligned decision

    [Zhouet al., 2024 ] Ruiwen Zhou, Yingxuan Yang, Muning Wen, Ying Wen, Wenhao Wang, Chunling Xi, Guoqiang Xu, Yong Yu, and Weinan Zhang. TRAD: Enhancing LLM agents with step-wise thought retrieval and aligned decision. InSIGIR, pages 3–13, 2024

    2024

  21. [2023]

    [Heet al., 2025 ] Gaole He, Gianluca Demartini, and Ujwal Gadiraju.Plan-Then-Execute: An Empirical Study of User Trust and Team Performance When Using LLM Agents As A Daily Assistant

    2025

  22. [2024]

    Memory in the age of AI agents.CoRR, abs/2512.13564,

    [Huet al., 2025 ] Yuyang Hu, Shichun Liu, Yanwei Yue, Guibin Zhang, Boyang Liu, Fangyi Zhu, Jiahang Lin, Honglin Guo, Shihan Dou, Zhiheng Xi, et al. Memory in the age of AI agents.CoRR, abs/2512.13564,

    Pith/arXiv arXiv 2025

  23. [2025]

    Mem0: Build- ing production-ready AI agents with scalable long-term memory.CoRR, abs/2504.19413,

    [Chhikaraet al., 2025 ] Prateek Chhikara, Dev Khant, Saket Aryan, Taranjeet Singh, and Deshraj Yadav. Mem0: Build- ing production-ready AI agents with scalable long-term memory.CoRR, abs/2504.19413,

    Pith/arXiv arXiv 2025

  24. [2026]

    G- Memory: Tracing hierarchical memory for multi-agent systems.CoRR, abs/2506.07398,

    [Zhanget al., 2025a ] Guibin Zhang, Muxin Fu, Guancheng Wan, Miao Yu, Kun Wang, and Shuicheng Yan. G- Memory: Tracing hierarchical memory for multi-agent systems.CoRR, abs/2506.07398,

    Pith/arXiv arXiv