pith. sign in

arxiv: 2607.01935 · v1 · pith:OZCA35CNnew · submitted 2026-07-02 · 💻 cs.AI

A-TMA: Decoupling State-Aware Memory Failures in Long-Term Agent Memory

Zitong Shi , Yixuan Tang , Anthony Kum Hoe Tung This is my paper

Pith reviewed 2026-07-03 14:00 UTC · model grok-4.3

classification 💻 cs.AI
keywords ghost memorylong-term agent memorystate-aware memoryLLM agentstemporal reasoningmemory failuresdecoupled evaluationLTP benchmark
0
0 comments X

The pith

Explicit state labels and evidence packets reduce ghost memory failures hidden by standard QA accuracy in long-term agent memory.

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

The paper establishes that ghost memory arises when superseded, current, and transition facts coexist in a memory bank, mix during retrieval, and mislead the answer model on changing user facts. It argues that memory systems must be optimized separately at the bank, retrieval, and answer-resolution levels rather than judged only by final accuracy. ATMA supplies a state-aware overlay that retains transition records, assembles evidence packets matched to the query's requested state view, and passes current, historical, and transition labels to the QA stage. The authors demonstrate that these additions improve conflict accuracy on a new benchmark and temporal F1 on an existing one, showing that state coordination failures can be isolated and reduced.

Core claim

Ghost memory is a state coordination failure in which old, current, and transition facts remain mixed in the memory bank and during retrieval, leading the answer model astray. Memory systems should be understood and optimized at three distinct levels: bank maintenance, retrieval, and answer-time resolution. ATMA is a state-aware overlay that keeps superseded and transition records, builds evidence packets for the query's requested state view, and exposes current, historical, and transition labels to QA. Because final QA accuracy conceals where these failures occur, the authors introduce decoupled evaluation and the LTP benchmark to measure them directly.

What carries the argument

ATMA, a state-aware overlay that retains superseded and transition records in the bank, constructs evidence packets aligned to a query's state view, and supplies current/historical/transition labels to the QA model.

If this is right

  • Keeping superseded and transition records prevents irreversible loss of change history in the memory bank.
  • State-specific evidence packets limit retrieval of facts incompatible with the query's requested time view.
  • Exposing state labels to QA lets the answer model resolve contradictions that would otherwise remain hidden.
  • Decoupled evaluation at bank, retrieval, and answer levels reveals the precise location of memory failures.
  • Improvements on LTP and LoCoMo indicate that state coordination can be added as an overlay to existing memory architectures.

Where Pith is reading between the lines

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

  • The host-dependent nature of the gains implies that integration details between ATMA and different base memory systems will determine how widely the approach transfers.
  • If state labels prove effective, future memory benchmarks may need to include explicit queries for current versus historical states rather than relying on overall accuracy.
  • The three-level decoupling framework could be applied to other persistent-agent components such as planning or tool-use memory to isolate similar coordination failures.

Load-bearing premise

That explicit state labels and evidence packets will measurably reduce ghost memory failures on conflict-heavy queries without creating new retrieval or answer errors that offset the gains.

What would settle it

Running Graphiti+ATMA and several other base memory systems on the LTP benchmark and finding no increase, or a decrease, in conflict accuracy while standard QA accuracy stays flat or falls.

Figures

Figures reproduced from arXiv: 2607.01935 by Anthony Kum Hoe Tung, Yixuan Tang, Zitong Shi.

Figure 1
Figure 1. Figure 1: Subset Statistics. Timestamped or graph la￾beled evidence can still fail at state correct QA. Explicit state roles reduce this failure in LTP and improve the Graphiti pair on LoCoMo sample0. The same observation creates an evaluation prob￾lem. A final answer can be wrong for different reasons. The memory bank may fail to preserve old and current states cleanly. Retrieval may choose evidence for the wrong s… view at source ↗
Figure 2
Figure 2. Figure 2: A-TMA Framework. Ghost memory mixes active, transition, and superseded facts across the bank, retrieval track, and generator. We separate this pipeline into bank state maintenance, state-aligned retrieval, and QA with explicit state roles, so each stage can be diagnosed and optimized before the final answer is produced. discuss the same state variable and whether their logical stances differ enough to need… view at source ↗
read the original abstract

Long term memory lets LLM agents act as persistent assistants, but user facts change. A useful memory system must know what is true now, what used to be true, and what changed. We study \emph{ghost memory}, a state coordination failure in which old, current, and transition facts coexist in the memory bank, remain mixed during retrieval, and mislead the answer model. We argue that memory systems should be understood and optimized from three levels: bank maintenance, retrieval, and answer time resolution. We propose ATMA, a state aware overlay for existing memory systems. ATMA keeps superseded and transition records in the bank, builds evidence packets for the query's requested state view, and exposes current, historical, and transition labels to QA. We further call for decoupled evaluation of bank, retrieval, and answer level failures, since final QA accuracy can hide where ghost memory occurs. To make this failure measurable, we build LTP (LoCoMo Temporal Plus), a conflict heavy benchmark for ghost memory, and evaluate on LoCoMo for long conversation generalization. On LTP, Graphiti+ATMA improves conflict accuracy by 0.240 absolute over Graphiti. On LoCoMo, Graphiti+ATMA raises temporal F1 from 0.0295 to 0.1705. The gains are host dependent, but they indicate that explicit state roles can reduce memory failures hidden by final QA accuracy.

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 defines 'ghost memory' as a state-coordination failure in long-term agent memory where superseded, current, and transition facts coexist and mix during retrieval, misleading the answer model. It proposes ATMA, a state-aware overlay for existing memory systems that retains superseded and transition records, constructs evidence packets matched to the query's requested state view, and supplies current/historical/transition labels to the QA stage. The work advocates decoupled evaluation at bank, retrieval, and answer levels (since aggregate QA accuracy can obscure failure locations), introduces the LTP benchmark for conflict-heavy temporal evaluation, and reports that Graphiti+ATMA improves conflict accuracy by 0.240 on LTP and temporal F1 from 0.0295 to 0.1705 on LoCoMo.

Significance. If the central claim is supported by the missing decoupled metrics and ablations, the work would supply a concrete mechanism and evaluation protocol for handling state evolution in persistent agent memory, moving beyond aggregate QA scores that the paper itself identifies as insufficient.

major comments (2)
  1. [Abstract] Abstract: the paper explicitly states that 'final QA accuracy can hide where ghost memory occurs' and calls for decoupled bank/retrieval/answer metrics, yet reports only end-to-end numbers (conflict accuracy +0.240 on LTP; temporal F1 +0.141 on LoCoMo) with no per-state retrieval precision, superseded-record pollution counts, or ablation isolating the label/packet component; this directly undermines the claim that state labels and evidence packets are responsible for the observed gains.
  2. [Abstract] Evaluation (LTP and LoCoMo results): no implementation details, controls, or error analysis are supplied for the reported numeric gains, making it impossible to determine whether the improvements arise from the claimed state coordination or from incidental changes in retrieval volume or host prompting.
minor comments (2)
  1. [Title] Title uses 'A-TMA' while the abstract and text consistently use 'ATMA'; standardize notation.
  2. [Abstract] The term 'ghost memory' is introduced without a formal definition or equation; a precise characterization (e.g., a set-equation over memory-bank states) would strengthen the subsequent claims.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for highlighting the gap between the paper's advocacy for decoupled evaluation and the reported results. We agree the claims would be stronger with explicit per-component metrics and controls. Below we respond to each major comment and commit to revisions that address them directly.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the paper explicitly states that 'final QA accuracy can hide where ghost memory occurs' and calls for decoupled bank/retrieval/answer metrics, yet reports only end-to-end numbers (conflict accuracy +0.240 on LTP; temporal F1 +0.141 on LoCoMo) with no per-state retrieval precision, superseded-record pollution counts, or ablation isolating the label/packet component; this directly undermines the claim that state labels and evidence packets are responsible for the observed gains.

    Authors: We acknowledge the inconsistency. Although the manuscript argues for bank/retrieval/answer decoupling and introduces LTP to surface ghost-memory failures, the presented results remain aggregate. In revision we will add (1) per-state retrieval precision, (2) superseded-record pollution counts at the bank level, and (3) an ablation that isolates the contribution of state labels plus evidence packets versus the host system alone. These additions will be placed in a new 'Decoupled Evaluation' subsection and will directly test whether the observed gains are attributable to the proposed mechanisms. revision: yes

  2. Referee: [Abstract] Evaluation (LTP and LoCoMo results): no implementation details, controls, or error analysis are supplied for the reported numeric gains, making it impossible to determine whether the improvements arise from the claimed state coordination or from incidental changes in retrieval volume or host prompting.

    Authors: The full manuscript contains an experimental-setup section, yet we agree it lacks the granularity needed to isolate causes. In the revision we will expand the evaluation to include: retrieval-volume-matched controls, host-prompting ablations, and a per-example error analysis that categorizes remaining failures by bank, retrieval, and answer stage. These controls will be reported alongside the existing LTP and LoCoMo numbers so readers can assess whether gains stem from state coordination. revision: yes

Circularity Check

0 steps flagged

No circularity in derivation chain

full rationale

The paper proposes the ATMA overlay system and reports empirical QA gains on LTP and LoCoMo benchmarks. No equations, fitted parameters, self-definitional constructs, or load-bearing self-citations appear in the provided text. The central claims rest on experimental results rather than reducing to inputs by construction, and the call for decoupled metrics is an evaluation recommendation, not a circular derivation step.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 1 invented entities

Abstract-only review; no explicit free parameters, axioms, or invented entities beyond the named concepts of ghost memory and ATMA can be extracted.

invented entities (1)
  • ghost memory no independent evidence
    purpose: Label for the state coordination failure where old, current, and transition facts coexist and mislead retrieval
    New term introduced in the abstract to describe the target failure mode.

pith-pipeline@v0.9.1-grok · 5790 in / 1134 out tokens · 34189 ms · 2026-07-03T14:00:17.602306+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

57 extracted references · 39 canonical work pages · 26 internal anchors

  1. [1]

    Generative Agents: Interactive Simulacra of Human Behavior

    Joon Sung Park, Joseph C. O’Brien, Carrie J. Cai, Meredith Ringel Morris, Percy Liang, and Michael S. Bernstein. Generative agents: Interactive simulacra of human behavior.arXiv 11 preprint arXiv:2304.03442, 2023. doi: 10.48550/arXiv.2304.03442. URL https://arxiv. org/abs/2304.03442

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2304.03442 2023

  2. [2]

    ReAct: Synergizing Reasoning and Acting in Language Models

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. ReAct: Synergizing reasoning and acting in language models.arXiv preprint arXiv:2210.03629,

    work page internal anchor Pith review Pith/arXiv arXiv

  3. [3]

    ReAct: Synergizing Reasoning and Acting in Language Models

    doi: 10.48550/arXiv.2210.03629. URLhttps://arxiv.org/abs/2210.03629

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2210.03629

  4. [4]

    Reflexion: Language Agents with Verbal Reinforcement Learning

    Noah Shinn, Federico Cassano, Edward Berman, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. Reflexion: Language agents with verbal reinforcement learning.arXiv preprint arXiv:2303.11366, 2023. doi: 10.48550/arXiv.2303.11366. URL https://arxiv.org/abs/ 2303.11366

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2303.11366 2023

  5. [5]

    Voyager: An Open-Ended Embodied Agent with Large Language Models

    Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. V oyager: An open-ended embodied agent with large language models.arXiv preprint arXiv:2305.16291, 2023. doi: 10.48550/arXiv.2305.16291. URL https://arxiv.org/abs/2305.16291

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2305.16291 2023

  6. [6]

    A Survey on Large Language Model based Autonomous Agents

    Lei Wang, Chen Ma, Xueyang Feng, Zeyu Zhang, Hao Yang, Jingsen Zhang, Zhiyuan Chen, Jiakai Tang, Xu Chen, Yankai Lin, Wayne Xin Zhao, Zhewei Wei, and Ji-Rong Wen. A survey on large language model based autonomous agents.arXiv preprint arXiv:2308.11432, 2023. doi: 10.48550/arXiv.2308.11432. URLhttps://arxiv.org/abs/2308.11432

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2308.11432 2023

  7. [7]

    Mem0: Building Production-Ready AI Agents with Scalable Long-Term Memory

    Prateek Chhikara, Dev Khant, Saket Aryan, Taranjeet Singh, and Deshraj Yadav. Mem0: Building production-ready AI agents with scalable long-term memory.arXiv preprint arXiv:2504.19413, 2025. doi: 10.48550/arXiv.2504.19413. URL https://arxiv.org/ abs/2504.19413

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2504.19413 2025

  8. [8]

    A-MEM: Agentic Memory for LLM Agents

    Wujiang Xu, Zujie Liang, Kai Mei, Hang Gao, Juntao Tan, and Yongfeng Zhang. A-Mem: Agentic memory for LLM agents. InAdvances in Neural Information Processing Systems, 2025. URLhttps://arxiv.org/abs/2502.12110. arXiv:2502.12110

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  9. [9]

    Zep: A Temporal Knowledge Graph Architecture for Agent Memory

    Preston Rasmussen, Pavlo Paliychuk, Travis Beauvais, Jack Ryan, and Daniel Chalef. Zep: A temporal knowledge graph architecture for agent memory, 2025. URL https://arxiv.org/ abs/2501.13956

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  10. [10]

    MemoryBank: Enhancing Large Language Models with Long-Term Memory

    Wanjun Zhong, Lianghong Guo, Qiqi Gao, He Ye, and Yanlin Wang. MemoryBank: Enhancing large language models with long-term memory.arXiv preprint arXiv:2305.10250, 2023. doi: 10.48550/arXiv.2305.10250. URLhttps://arxiv.org/abs/2305.10250

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2305.10250 2023

  11. [11]

    MemGPT: Towards LLMs as Operating Systems

    Charles Packer, Sarah Wooders, Kevin Lin, Vivian Fang, Shishir G. Patil, Ion Stoica, and Joseph E. Gonzalez. MemGPT: Towards LLMs as operating systems.arXiv preprint arXiv:2310.08560, 2023. doi: 10.48550/arXiv.2310.08560. URL https://arxiv.org/ abs/2310.08560

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2310.08560 2023

  12. [12]

    MEMORYLLM: Towards self-updatable large language models.arXiv preprint arXiv:2402.04624, 2024

    Yu Wang, Yifan Gao, Xiusi Chen, Haoming Jiang, Shiyang Li, Jingfeng Yang, Qingyu Yin, Zheng Li, Xian Li, Bing Yin, Jingbo Shang, and Julian McAuley. MEMORYLLM: Towards self-updatable large language models.arXiv preprint arXiv:2402.04624, 2024. doi: 10.48550/ arXiv.2402.04624. URLhttps://arxiv.org/abs/2402.04624

    work page arXiv 2024

  13. [13]

    MemOS: A Memory OS for AI System

    Zhiyu Li, Chenyang Xi, Chunyu Li, Ding Chen, Boyu Chen, Shichao Song, Simin Niu, Hanyu Wang, Jiawei Yang, Chen Tang, Qingchen Yu, Jihao Zhao, et al. MemOS: A memory OS for AI system.arXiv preprint arXiv:2507.03724, 2025. doi: 10.48550/arXiv.2507.03724. URL https://arxiv.org/abs/2507.03724

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2507.03724 2025

  14. [14]

    Vicky Zhao, Lili Qiu, and Jianfeng Gao

    Zhuoshi Pan, Qianhui Wu, Huiqiang Jiang, Xufang Luo, Hao Cheng, Dongsheng Li, Yuqing Yang, Chin-Yew Lin, H. Vicky Zhao, Lili Qiu, and Jianfeng Gao. On memory construction and retrieval for personalized conversational agents.arXiv preprint arXiv:2502.05589, 2025. doi: 10.48550/arXiv.2502.05589. URLhttps://arxiv.org/abs/2502.05589

    work page doi:10.48550/arxiv.2502.05589 2025

  15. [15]

    Augmenting language models with long-term memory.arXiv preprint arXiv:2306.07174, 2023

    Weizhi Wang, Li Dong, Hao Cheng, Xiaodong Liu, Xifeng Yan, Jianfeng Gao, and Furu Wei. Augmenting language models with long-term memory.arXiv preprint arXiv:2306.07174, 2023. doi: 10.48550/arXiv.2306.07174. URLhttps://arxiv.org/abs/2306.07174. 12

    work page doi:10.48550/arxiv.2306.07174 2023

  16. [16]

    M+: Extending MemoryLLM with scalable long-term memory.arXiv preprint arXiv:2502.00592, 2025

    Yu Wang, Dmitry Krotov, Yuanzhe Hu, Yifan Gao, Wangchunshu Zhou, Julian McAuley, Dan Gutfreund, Rogerio Feris, and Zexue He. M+: Extending MemoryLLM with scalable long-term memory.arXiv preprint arXiv:2502.00592, 2025. doi: 10.48550/arXiv.2502.00592. URL https://arxiv.org/abs/2502.00592

    work page doi:10.48550/arxiv.2502.00592 2025

  17. [17]

    AriadneMem: Threading the maze of lifelong memory for LLM agents, 2026

    Wenhui Zhu, Xiwen Chen, Zhipeng Wang, Jingjing Wang, Xuanzhao Dong, Minzhou Huang, Rui Cai, Hejian Sang, Hao Wang, Peijie Qiu, Yueyue Deng, Prayag Tiwari, Brendan Hogan Rappazzo, and Yalin Wang. AriadneMem: Threading the maze of lifelong memory for LLM agents, 2026. URLhttps://arxiv.org/abs/2603.03290

    work page arXiv 2026

  18. [18]

    HippoRAG: Neurobiologically inspired long-term memory for large language models.arXiv preprint arXiv:2405.14831, 2024

    Bernal Jimenez Gutierrez, Yiheng Shu, Yu Gu, Michihiro Yasunaga, and Yu Su. HippoRAG: Neurobiologically inspired long-term memory for large language models.arXiv preprint arXiv:2405.14831, 2024. doi: 10.48550/arXiv.2405.14831. URL https://arxiv.org/abs/ 2405.14831

    work page doi:10.48550/arxiv.2405.14831 2024

  19. [19]

    AtomMem: Learnable dynamic agentic memory with atomic memory operation.arXiv preprint arXiv:2601.08323,

    Yupeng Huo, Yaxi Lu, Zhong Zhang, Haotian Chen, and Yankai Lin. AtomMem: Learnable dynamic agentic memory with atomic memory operation.arXiv preprint arXiv:2601.08323,

    work page arXiv

  20. [20]

    AtomMem: Learnable dynamic agentic memory with atomic memory operation.arXiv preprint arXiv:2601.08323,

    doi: 10.48550/arXiv.2601.08323. URLhttps://arxiv.org/abs/2601.08323

    work page doi:10.48550/arxiv.2601.08323

  21. [21]

    Graphiti: Build temporal context graphs for ai agents

    Zep AI. Graphiti: Build temporal context graphs for ai agents. https://github.com/ getzep/graphiti, 2024. Open-source temporal context graph engine underlying Zep, ac- cessed 2026-06-06

    2024

  22. [22]

    Evaluating very long-term conversational memory of LLM agents

    Adyasha Maharana, Dong-Ho Lee, Sergey Tulyakov, Mohit Bansal, Francesco Barbieri, and Yuwei Fang. Evaluating very long-term conversational memory of LLM agents. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (V olume 1: Long Papers), pages 13851–13870, Bangkok, Thailand, August 2024. Association for Computational...

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

  23. [23]

    $\delta$-mem: Efficient Online Memory for Large Language Models

    Jingdi Lei, Di Zhang, Junxian Li, Weida Wang, Kaixuan Fan, Xiang Liu, Qihan Liu, Xiaoteng Ma, Baian Chen, and Soujanya Poria. δ-mem: Efficient online memory for large language models.arXiv preprint arXiv:2605.12357, 2026. doi: 10.48550/arXiv.2605.12357. URL https://arxiv.org/abs/2605.12357

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2605.12357 2026

  24. [24]

    MemTrace: Tracing and Attributing Errors in Large Language Model Memory Systems

    Xinle Deng, Ruobin Zhong, Hujin Peng, Xiaoben Lu, Yanzhe Wu, Guang Li, Buqiang Xu, Yunzhi Yao, Jizhan Fang, Haoliang Cao, Junjie Guo, Yuan Yuan, Ziqing Ma, Yuanqiang Yu, Rui Hu, Baohua Dong, Hangcheng Zhu, and Ningyu Zhang. MemTrace: Tracing and attributing errors in large language model memory systems.arXiv preprint arXiv:2605.28732, 2026. doi: 10.48550/...

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2605.28732 2026

  25. [25]

    MemTrace: Probing What Final Accuracy Misses in Long-Term Memory

    Xianxuan Long, Zhikai Chen, Shenglai Zeng, Shouren Wang, Kai Guo, and Jiliang Tang. MemTrace: Probing what final accuracy misses in long-term memory.arXiv preprint arXiv:2606.17328, 2026. URLhttps://arxiv.org/abs/2606.17328

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  26. [26]

    DynamicMem: A Long-Horizon Memory Benchmark in Real-World Settings

    Wenya Xie, Shengming Zhou, Zelin Li, Pouya Parsa, Shuang Zhou, Xinheng Ding, Chin- may Arvind, Guanchu Wang, Vladimir Braverman, Ali Payani, Yantao Zheng, and Zirui Liu. DynamicMem: A long-horizon memory benchmark in real-world settings.arXiv preprint arXiv:2606.22877, 2026. doi: 10.48550/arXiv.2606.22877. URL https://arxiv.org/abs/ 2606.22877

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2606.22877 2026

  27. [27]

    MEMPROBE: Probing Long-Term Agent Memory via Hidden User-State Recovery

    Enze Ma, Yufan Zhou, Wei-Chieh Huang, Jie Yang, Huanhuan Ma, Zixuan Wang, Chengze Li, Chunyu Miao, Philip S. Yu, and Zhen Wang. MEMPROBE: Probing long-term agent memory via hidden user-state recovery.arXiv preprint arXiv:2606.24595, 2026. doi: 10.48550/arXiv. 2606.24595. URLhttps://arxiv.org/abs/2606.24595

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv 2026

  28. [28]

    Are We Ready For An Agent-Native Memory System?

    Wei Zhou, Xuanhe Zhou, Shaokun Han, Hongming Xu, Guoliang Li, Zhiyu Li, Feiyu Xiong, and Fan Wu. Are we ready for an agent-native memory system?arXiv preprint arXiv:2606.24775,

    work page internal anchor Pith review Pith/arXiv arXiv

  29. [29]

    Are We Ready For An Agent-Native Memory System?

    doi: 10.48550/arXiv.2606.24775. URLhttps://arxiv.org/abs/2606.24775

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2606.24775

  30. [30]

    LongBench: A Bilingual, Multitask Benchmark for Long Context Understanding

    Yushi Bai, Xin Lv, Jiajie Zhang, Hongchang Lyu, Jiankai Tang, Zhidian Huang, Zhengxiao Du, Xiao Liu, Aohan Zeng, Lei Hou, Yuxiao Dong, Jie Tang, and Juanzi Li. LongBench: A bilingual, multitask benchmark for long context understanding.arXiv preprint arXiv:2308.14508, 2023. doi: 10.48550/arXiv.2308.14508. URLhttps://arxiv.org/abs/2308.14508. 13

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2308.14508 2023

  31. [31]

    LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory

    Di Wu, Hongwei Wang, Wenhao Yu, Yuwei Zhang, Kai-Wei Chang, and Dong Yu. Long- MemEval: Benchmarking chat assistants on long-term interactive memory.arXiv preprint arXiv:2410.10813, 2024. doi: 10.48550/arXiv.2410.10813. URL https://arxiv.org/abs/ 2410.10813

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2410.10813 2024

  32. [32]

    Lost in the Middle: How Language Models Use Long Contexts

    Nelson F. Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. Lost in the middle: How language models use long contexts.arXiv preprint arXiv:2307.03172, 2023. doi: 10.48550/arXiv.2307.03172. URL https://arxiv.org/abs/ 2307.03172

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2307.03172 2023

  33. [33]

    REALM: Retrieval-Augmented Language Model Pre-Training

    Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat, and Ming-Wei Chang. REALM: Retrieval-augmented language model pre-training.arXiv preprint arXiv:2002.08909, 2020. doi: 10.48550/arXiv.2002.08909. URLhttps://arxiv.org/abs/2002.08909

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2002.08909 2002

  34. [34]

    Dense Passage Retrieval for Open-Domain Question Answering

    Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. Dense passage retrieval for open-domain question answering. arXiv preprint arXiv:2004.04906, 2020. doi: 10.48550/arXiv.2004.04906. URL https: //arxiv.org/abs/2004.04906

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2004.04906 2004

  35. [35]

    Retrieval-augmented generation for knowledge-intensive NLP tasks

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Na- man Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, and Douwe Kiela. Retrieval-augmented generation for knowledge-intensive NLP tasks. InAdvances in Neural Information Processing Systems, volume 33, pages 9459–9474, 2020. URL https://...

    2020

  36. [36]

    Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection

    Akari Asai, Zeqiu Wu, Yizhong Wang, Avirup Sil, and Hannaneh Hajishirzi. Self-RAG: Learn- ing to retrieve, generate, and critique through self-reflection.arXiv preprint arXiv:2310.11511,

    work page internal anchor Pith review Pith/arXiv arXiv

  37. [37]

    Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection

    doi: 10.48550/arXiv.2310.11511. URLhttps://arxiv.org/abs/2310.11511

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2310.11511

  38. [38]

    Chain-of-note: Enhancing robustness in retrieval-augmented language models.arXiv preprint arXiv:2311.09210, 2023

    Wenhao Yu, Hongming Zhang, Xiaoman Pan, Kaixin Ma, Hongwei Wang, and Dong Yu. Chain-of-note: Enhancing robustness in retrieval-augmented language models.arXiv preprint arXiv:2311.09210, 2023. doi: 10.48550/arXiv.2311.09210. URL https://arxiv.org/abs/ 2311.09210

    work page doi:10.48550/arxiv.2311.09210 2023

  39. [39]

    Sentence-BERT: Sentence embeddings using siamese BERT-networks

    Nils Reimers and Iryna Gurevych. Sentence-BERT: Sentence embeddings using siamese BERT-networks. InProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing, pages 3982–3992, 2019. doi: 10.18653/v1/D19-1410. URL https: //aclanthology.org/D19-1410/

    work page doi:10.18653/v1/d19-1410 2019

  40. [40]

    Decoupled weight decay regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. InInternational Conference on Learning Representations, 2019. URL https://openreview.net/forum? id=Bkg6RiCqY7

    2019

  41. [41]

    Inside out: Evolving user-centric core memory trees for long-term personalized dialogue systems.arXiv preprint arXiv:2601.05171, 2026

    Jihao Zhao, Ding Chen, Zhaoxin Fan, Kerun Xu, Mengting Hu, Bo Tang, Feiyu Xiong, and Zhiyu Li. Inside out: Evolving user-centric core memory trees for long-term personalized dialogue systems.arXiv preprint arXiv:2601.05171, 2026. doi: 10.48550/arXiv.2601.05171. URLhttps://arxiv.org/abs/2601.05171

    work page doi:10.48550/arxiv.2601.05171 2026

  42. [42]

    B leu: a Method for Automatic Evaluation of Machine Translation

    Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. BLEU: A method for automatic evaluation of machine translation. InProceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311–318, 2002. doi: 10.3115/1073083.1073135. URL https://aclanthology.org/P02-1040/. A Experimental Details A.1 Benchmark Protocol All ...

    work page doi:10.3115/1073083.1073135 2002

  43. [43]

    infer the requested state view: current, historical, change, or neutral

  44. [44]

    select and rank only the highest-priority candidate evidence rows for state-correct evidence construction

  45. [45]

    prefer direct state evidence over broad weak hops

  46. [46]

    keep transition evidence only when it helps answer a change-related or ambiguity-sensitive query

  47. [47]

    be concise: short reasons only

  48. [48]

    Before the update, how many hours per night did you sleep on average?

    return strict JSON only. The user prompt is a JSON object with task=retrieve_control, the query, top_k, max_selected, the rule based query profile, the compact candidate list, the required output schema, and short instructions. The required output fields are query_mode, confidence, fallback_recommended, a selected list containing candidate_id, rank, role,...

  49. [49]

    Speaker Caroline says : I sleep about 7 hours per night on average

    “Speaker Caroline says : I sleep about 7 hours per night on average.”

  50. [50]

    Speaker Caroline says : I sleep about 5 hours per night on average

    “Speaker Caroline says : I sleep about 5 hours per night on average”

  51. [51]

    Speaker Caroline says : Wow! Did you see that band?

    “Speaker Caroline says : Wow! Did you see that band?”

  52. [52]

    Speaker Melanie says : Wow, looks awesome! Did you join in?

    “Speaker Melanie says : Wow, looks awesome! Did you join in?” QA serialization.Full A-TMA passes the same evidence with the requested historical state view, while the no-label variant passes the rows without explicit state roles. QA responses.Full A-TMA: “Before the update, Caroline slept about 5 hours per night on average.” Judge: correct, 1.0. No QA lab...

  53. [56]

    Speaker Caroline says : I drive about 4500 miles per year in my 14-year-old car

    “Speaker Caroline says : I drive about 4500 miles per year in my 14-year-old car.” Retrieved evidence with controller (top rows)

  54. [57]

    Speaker Caroline says : I drive about 3000 miles per year in my 14-year-old car

    “Speaker Caroline says : I drive about 3000 miles per year in my 14-year-old car”

  55. [58]

    Speaker Caroline says : Sounds fun! What was the best part? Do you do it often with the kids?

    “Speaker Caroline says : Sounds fun! What was the best part? Do you do it often with the kids?”

  56. [59]

    Speaker Melanie says : Wow, sounds amazing! What was the event like? Those posters are great!

    “Speaker Melanie says : Wow, sounds amazing! What was the event like? Those posters are great!”

  57. [60]

    Speaker Caroline says : I drive about 4500 miles per year in my 14-year-old car

    “Speaker Caroline says : I drive about 4500 miles per year in my 14-year-old car.” QA responses.Shadow baseline: “4500 miles per year.” Judge: wrong, 0.0. Online controller: “3000 miles per year.” Judge: correct, 1.0. Run note.QA model:qwen2.5:3b; judge:qwen3:30b. The first trace isolates answer-side serialization: the retrieved rows are unchanged, but ex...