Pith. sign in

REVIEW 5 major objections 6 minor 74 references

TrajWiki: Source-Grounded Memory Trajectories for Long-Horizon Dialogue Agents

T0 review · 5 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Trajectory memory beats flat memory in long-horizon dialogue agents

desk verdict TrajWiki is a genuinely novel memory architecture with a real design contribution, but the headline claim of across-backbone improvement is not supported by its own Table 1 for Qwen3-8B; worth serious review with major revisions requested. read the letter →

arxiv 2608.00967 v1 pith:K4SUZNH3 submitted 2026-08-02 cs.AI

Jingyu Sun , Yuyang Xue , Mingyang Li , Zhengtao Yao , Jiachen Li , Yang Cui , Wenhao Cai , Haozhe Liu
show 4 more authors
Fangying Wang Magdalene Katharina Montgomery Syed Murtuza Baker Hongpeng Zhou
This is my paper
classification cs.AI
keywords long-horizondialoguememorytrajectoriessourcegroundingwikiprovenancehierarchicalretrievalLLMagentsretrieval-augmentedgeneration
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper argues that long-horizon dialogue agents fail not because they lack memory but because they store memory as flat, overwritable records that lose how facts originated and changed. It proposes TrajWiki, which represents each memory as a source-grounded evolution trajectory built from immutable snapshots and claim-level edits, and organizes these trajectories into a persistent, interlinked Memory Wiki. At query time, retrieval routes hierarchically through wiki pages, trajectories, and snapshots back to raw messages, so answers are grounded in source evidence. The paper reports that on LoCoMo and MedMT-Bench this design improves accuracy over flat-memory baselines and makes memory evolution and failures diagnosable.

What carries the argument

The core mechanism is the memory trajectory $T_j = (\sigma_{j,1}, \ldots, \sigma_{j,n_j})$, an append-only list of immutable episodic snapshots linked to raw messages, updated through claim-level operations such as ADD, REVISE, and DEPRECATE, together with the Memory Wiki—an intermediate layer of index, entity, topic, and inventory pages that routes queries to a smaller candidate set of trajectories before evidence expansion. The trajectory preserves provenance and update history; the wiki makes retrieval scalable and keeps related memories organized.

What would settle it

Run a large sample of queries where the gold evidence is present in the raw dialogue, and measure whether the fraction of gold facts that survive into the wiki page text predicts answer failure. If the abstraction-loss rate is high and correlates with failure, the claim that the trajectory-plus-wiki organization preserves source grounding is falsified. Alternatively, modify the pipeline to verify final answers directly against raw source messages rather than summaries; if that verification closes most answer gaps, the abstraction loss is the bottleneck.

Watch

Extended reading notes

Core claim

The central discovery is that representing conversational memory as an append-only sequence of immutable, source-linked episodic snapshots—with claim-level ADD, REVISE, and DEPRECATE operations—and compiling those trajectories into a wiki-style intermediate layer yields better long-horizon dialogue answers than retrieving isolated memory records. The trajectory preserves how a belief evolved, the wiki reduces fragmentation, and hierarchical retrieval lets queries move from wiki pages to trajectories to raw source messages. The paper shows gains across open- and closed-source backbones, with the largest improvements on multi-hop questions and on a medical benchmark's contradiction subset.

Load-bearing premise

The framework assumes that the LLM-based steps of claim extraction, trajectory matching, and wiki compilation preserve the specific source facts well enough that later answer synthesis can still recover them; the paper's own case study (query conv-26_qa_24) shows a pottery fact present in raw evidence was lost through abstraction, so this assumption can fail.

Editorial extensions

If this is right

  • Multi-hop questions benefit most: answers requiring cross-session aggregation improve because complementary trajectories are retrieved together.
  • Memory updates become auditable: no claim is deleted, so revision and deprecation history can be inspected.
  • The Memory Wiki shrinks the query-time search space (a 2.35x reduction in the paper), helping retrieval scale as memory grows.
  • Answers can be traced to source messages, so unsupported evidence and failure stages can be localized and diagnosed.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: the same trajectory-plus-wiki organization could extend to non-dialogue agent memories, such as task logs or tool-use histories, wherever provenance and update history matter.
  • The documented abstraction-loss case suggests a testable extension: a consistency audit comparing wiki-page claims against raw source messages before answer generation would catch salience drift.
  • The cost profile (memory construction dominating) implies the framework is most economical when memory is reused across many queries, a point the paper notes but does not quantify as a break-even.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 6 minor

Summary. The paper proposes TrajWiki, a long-horizon dialogue memory framework that represents memories as source-grounded, immutable episodic snapshot trajectories with claim-level ADD/REVISE/DEPRECATE operations, organizes trajectories into a persistent Memory Wiki with index/entity/topic/inventory pages, and performs hierarchical retrieval (wiki pages -> trajectories -> snapshots -> source messages) before evidence-grounded answer synthesis. The method is evaluated on LoCoMo and MedMT-Bench with GPT-4o-mini, Qwen3-32B, and Qwen3-8B backbones against Full Context, Naive RAG, LangMem, A-MEM, and Mem0, reporting F1, BLEU, LLM-judge accuracy, and rubric-based pass rates, along with hyperparameter analyses, counterfactual ablations, cost/scalability breakdowns, and diagnostic failure analyses. The central claim is that TrajWiki improves long-horizon dialogue performance across both open-source and closed-source LLM backbones while adding interpretability and diagnostic visibility.

Significance. If the central claim held, TrajWiki would be a meaningful step beyond flat memory retrieval: the trajectory-plus-wiki organization is well motivated, and the paper provides an unusually high level of implementation detail, including deterministic scoring features, structured-output schemas, fallback and repair procedures, prompts, cost accounting, and explicit diagnostic limitations. The appendices are a genuine strength: they make the pipeline reproducible and honestly document failure modes, including the pottery-fact salience-loss case. However, the empirical support is currently mixed. The abstract's across-backbone claim is contradicted by the Qwen3-8B F1 results in Table 1, and the evaluation relies on author-defined canonicalized metrics and a permissive LLM judge without human calibration. These issues are load-bearing for the paper's main empirical assertion, so the manuscript needs substantial revision before the claimed contribution can be accepted.

major comments (5)
  1. [Abstract; Section 4.3; Table 1] The claim that TrajWiki 'improves long-horizon dialogue performance across both open-source and closed-source LLM backbones' is not supported by the open-source results. For Qwen3-8B, TrajWiki's multi-hop F1 is 30.25, below Mem0's 31.73; temporal F1 is 31.48, below Full Context's 33.64; and open-domain F1 is 16.89, below Full Context's 20.67. The gains in Section 4.3 are stated as 'most pronounced on multi-hop questions,' but for Qwen3-8B the multi-hop improvement appears only in Acc/BLEU, not F1. No error bars, significance tests, or pre-specified primary metric are given, so it is not possible to tell whether the apparent across-backbone success is driven by selected cells. The abstract and Section 4.3 need qualification, or the authors need to report a fully specified primary analysis with appropriate statistical support.
  2. [Appendix B.1; Section 4.1] The headline metrics F1 and BLEU-1 are not standard LoCoMo metrics: they are computed over a canonicalized semantic slot representation with a custom soft-matching policy that treats e.g. 'LGBTQ support group' and 'support group' as matches. No human calibration, inter-annotator agreement, or comparison against the original LoCoMo evaluation script is reported. The LLM judge in Appendix C.3 is also permissive by design (e.g., treating source-backed aliases as covered and allowing 'harmless same-category extras'). Because the central empirical claim depends on these metrics, the paper should either use the benchmarks' standard scoring or provide evidence that the custom metrics agree with human judgments and do not systematically favor TrajWiki.
  3. [Section 4.1] The decision to exclude the LoCoMo adversarial question category is a post-hoc benchmark modification. The rationale 'Following Mem0' is not a methodological justification, and it changes the task distribution against which the method is evaluated. The paper should report results on the full LoCoMo QA set or on the adversarial split separately, and should justify why excluding a category that tests exactly the kind of conflict/update handling TrajWiki claims to support is appropriate for the paper's central claim.
  4. [Table 1; Section 4.3] No measure of variability is reported for any experiment. Many of the claimed wins are small (e.g., GPT-4o-mini multi-hop F1: 35.23 vs. Mem0's 34.52; Qwen3-32B multi-hop Acc: 50.33 vs. 45.74), and the Table reports a single run per condition. Without multiple runs, bootstrap intervals, or significance tests, the reader cannot assess whether the differences are stable. This is particularly important because the paper's central claim is comparative.
  5. [Section 4.5; Appendix B.2] The counterfactual ablation is informative as a retrieval-proxy analysis, but it does not rerun answer generation for the ablated variants, so it cannot directly support the claim that the full TrajWiki organization improves final answer quality. The paper acknowledges this, but the abstract and Section 4.3 should be careful not to present the ablation as end-to-end causal evidence. A small end-to-end ablation (even on a subset) would substantially strengthen the argument.
minor comments (6)
  1. [Abstract] The benchmark name 'MedMT' appears in the abstract but is introduced as 'MedMT-Bench' in Section 4.1. Please use the full name consistently.
  2. [Table 1] Table 1 contains formatting errors, e.g., '44.7337.86' and '32.7179.68' without separators. Please fix the table formatting and ensure all numbers are legible.
  3. [Section 4.3] The phrase 'consistently outperforms most of the baselines' is vague. Please specify which baselines are outperformed on which metric and split, especially given the Qwen3-8B F1 exceptions noted above.
  4. [Appendix B.1] The definition of BLEU-1 drops the brevity penalty, which means it is not BLEU-1 in the standard sense. Please rename it (e.g., 'canonical unigram precision') or justify the deviation.
  5. [Appendix B.2 / Table 3] The abbreviation 'R@15' is used in Table 3 without a definition; it is only defined later in the text. Please define it at first use.
  6. [Section 4.8] The limitations discussion is honest and welcome. The case study in Appendix B.4 is a useful concrete illustration, but it also undercuts the 'source-grounded' claim; consider moving a short version of this case into the main text so readers do not miss it.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: TrajWiki is an empirical system paper; its claimed improvements are benchmark measurements, not derivations from the method's own definitions.

full rationale

TrajWiki does not present a first-principles derivation; its central claim is an empirical comparison (Section 4.3, Table 1). I checked for the circularity patterns: (1) No self-definitional reduction: the trajectory/wiki pipeline (Eqs. 2–12) is a construction, not a derivation, and memory trajectories are not defined in terms of benchmark outcomes. (2) No fitted input called prediction: the hyperparameters m=15, t=15, k=15 are described in Section 4.2 and analyzed offline in Section 4.4 using gold coverage diagnostics, but the reported answer-level numbers are not a statistical transform of those gold labels; at most this is a test-set-tuning/overfitting concern, not a constructional equivalence. (3) No load-bearing self-citations: the reference list contains no prior work by the present authors; baselines are external works (Mem0, A-MEM, LangMem, Naive RAG). (4) No imported uniqueness theorem. (5) No ansatz smuggled in via citation: the hierarchical retrieval and wiki compilation are the paper's own design choices, not justified by a cited prior adoption. (6) No renaming of a known result: TrajWiki is presented as a system with components, and whether it outperforms baselines is an empirical question. The caveats in Section 4.8 and Appendix B.4 (e.g., the conv-26_qa_24 pottery fact losing salience through abstraction) are limitations, not circularity. The Qwen3-8B results contradict the abstract's across-backbone claim on some F1 splits, but that is an evidentiary/correctness issue, not a circularity issue. Therefore the paper is not circular.

Assumptions & free parameters 5 free parameters · 3 assumptions · 2 invented entities

The central empirical claim rests on hand-tuned retrieval coefficients, LLM extraction reliability, and benchmark choices, but does not introduce a physical entity or a derivation. The most important unverified dependency is that repeated LLM abstraction preserves source facts; the paper itself documents a counterexample.

free parameters (5)
  • Maximum trajectory length m = 15
    Set after offline snapshot-depth analysis on LoCoMo (Section 4.4, Figure 4); not validated on a held-out split.
  • Routed wiki pages t = 15
    Chosen from coverage-vs-cost analysis on LoCoMo (Figure 3); not a predicted value.
  • Selected trajectories k = 15
    Chosen from top-k coverage analysis on LoCoMo (Figure 3); affects recall and answer noise.
  • Trajectory continuation threshold delta = 0.72
    Deterministic fallback threshold in mock mode (Appendix A.2, A.7); no independent justification.
  • Retrieval scoring weights and bonuses = e.g. 0.60, 0.20, 0.20 in g_det; 0.75, 0.15, 0.10 in s_dense; bonuses 0.08, 0.06, 0.03; RRF constant 60
    Hand-tuned coefficients in Appendix A.2 and A.5; no search procedure or ablation of these weights is reported.
assumptions (3)
  • domain assumption LLM-based claim extraction, trajectory matching, and wiki compilation preserve source facts well enough for later answer synthesis.
    The whole pipeline depends on structured LLM outputs. Section 4.8 lists fallback and repair needs, and Appendix B.4 gives a concrete salience-loss failure, so this is load-bearing and imperfect.
  • domain assumption The custom canonicalized F1, BLEU, and LLM-as-a-judge protocol measure answer correctness in long-horizon dialogue.
    Appendix B.1 and C.3 define a bespoke soft-matching metric and a permissive equivalence policy, but no human calibration or inter-annotator agreement is reported.
  • ad hoc to paper Excluding the LoCoMo adversarial question category is appropriate for comparing against prior memory baselines.
    Section 4.1 drops the adversarial category 'to ensure fair comparison' following Mem0. This is a post-hoc benchmark choice that changes the aggregate difficulty.
invented entities (2)
  • Memory trajectory (ordered append-only snapshot sequence)
    purpose: Represent each memory as a source-grounded evolution history using claim-level ADD, REVISE, and DEPRECATE operations.
    Software construct in Equations (3) to (5); only evaluated through end-to-end system performance, no outside-paper falsifiable handle.
  • Memory Wiki page layer (index, entity, topic, inventory pages)
    purpose: Intermediate persistent organization layer that compiles trajectories into interlinked pages for hierarchical retrieval.
    Internal architectural component; its benefits are measured only through the authors' offline ablations and gold-evidence labels in Tables 2 to 4.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TrajWiki: Source-Grounded Memory Trajectories for Long-Horizon Dialogue Agents." pith.science (2026). https://pith.science/paper/K4SUZNH3

@misc{pith2026260800967,
  author       = {Pith},
  title        = {Pith review of: TrajWiki: Source-Grounded Memory Trajectories for Long-Horizon Dialogue Agents},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/K4SUZNH3}},
  note         = {Machine review of arXiv:2608.00967}
}
read the original abstract

Large language model agents have shown strong capabilities in generating coherent and contextually appropriate responses, yet robust long-horizon dialogue remains limited by the lack of external memory that is traceable, updatable, and diagnostically transparent. Existing memory-augmented agents often store memories as isolated records or overwritable states, making it difficult to preserve how information originates, evolves, conflicts, or becomes obsolete over time. We propose TrajWiki, a trajectory-based memory framework for long-horizon conversational agents. Instead of treating memory as static entries, TrajWiki represents each memory as a source-grounded evolution trajectory, maintained through immutable episodic snapshots and claim-level operations such as ADD, REVISE, and DEPRECATE. To reduce fragmentation and retrieval cost, TrajWiki further introduces Memory Wiki, a persistent intermediate layer that incrementally compiles dialogue history into structured and interlinked wiki pages capturing salient entities, events, quantities, topics, and conflicts. At inference time, queries are routed hierarchically from relevant wiki pages to linked memory trajectories, then to corresponding snapshots and source messages for evidence-grounded answer synthesis. Experiments on LoCoMo and MedMT show that TrajWiki improves long-horizon dialogue performance across both open-source and closed-source LLM backbones, while providing greater interpretability and diagnostic visibility into memory evolution, retrieval failures, and answer generation.

Figures

Figures reproduced from arXiv: 2608.00967 by the authors.

Figure 1
Figure 1. Comparison of conventional agent memory and TrajWiki’s source-grounded trajec [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of the TrajWiki architecture. TrajWiki converts dialogue history into source￾grounded memory trajectories, compiles them into a persistent and interlinked Memory Wiki, retrieves evidence hierarchically from wiki pages to trajectories and source messages, and generates answers grounded in the retrieved evidence. The Memory Wiki serves as a compounding knowledge artifact, maintaining cross-references, flagged… view at source ↗
Figure 3
Figure 3. Hyperparameter Retrieval Cutoff Analysis [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Snapshot Budget Analysis for trajectory length m on the LoCoMo. The left plot shows the cumulative fraction of queries whose gold evidence is covered within a given required snapshot rank. The right plot shows the marginal number of newly covered queries at each rank, …
Figure 5
Figure 5. Figure 5: Trajectory Analysis. The left panel shows head-tail semantic similarity for non-singleton trajectories, while the right panel breaks semantic drift down by trajectory length and contrasts global head-tail span with local adjacent-snapshot similarity. Most trajectories …
Figure 6
Figure 6. Figure 6: Token usage by cost phase. Token usage is dominated by memory construction and query-time retrieval, while benchmark-only evaluation contributes less than 1% of total measured tokens. 3409.2 claims, and 42.5 wiki pages [PITH_FULL_IMAGE:figures/full_fig_p019_6.png]
Figure 7
Figure 7. Figure 7: Provider-side latency by cost phase. Accumulated provider latency is primarily spent on memory construction, including extraction, trajectory organization, claim processing, and wiki compilation [PITH_FULL_IMAGE:figures/full_fig_p020_7.png]
Figure 8
Figure 8. Figure 8: Memory size across dialogue samples. Counts are shown on a log scale. TrajWiki stores raw messages, trajectories, snapshots, claims, and wiki pages as separate memory objects [PITH_FULL_IMAGE:figures/full_fig_p021_8.png]
Figure 9
Figure 9. Figure 9: Candidate universe scaling. Wiki routing keeps the trajectory candidate universe substantially smaller than direct retrieval over all stored trajectories. 21 [PITH_FULL_IMAGE:figures/full_fig_p021_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

74 extracted references · 45 canonical work pages

  1. [1]

    From language to action: a review of large language models as autonomous agents and tool users.Artificial Intelligence Review, 2026

    Sadia Sultana Chowa, Riasad Alvi, Subhey Sadi Rahman, Md Abdur Rahman, Mohaimenul Azam Khan Raiaan, Md Rafiqul Islam, Mukhtar Hussain, and Sami Azam. From language to action: a review of large language models as autonomous agents and tool users.Artificial Intelligence Review, 2026

    work page 2026

  2. [2]

    React: Synergizing reasoning and acting in language models.arXiv preprint arXiv:2210.03629, 2022

    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, 2022

    arXiv 2022

  3. [3]

    Toolformer: Language models can teach themselves to use tools.Advances in neural information processing systems, 36:68539–68551, 2023

    Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, Roberta Raileanu, Maria Lomeli, Eric Hambro, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. Toolformer: Language models can teach themselves to use tools.Advances in neural information processing systems, 36:68539–68551, 2023

    2023

  4. [4]

    V oyager: An open-ended embodied agent with large language models.arXiv preprint arXiv:2305.16291, 2023

    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

    arXiv 2023

  5. [5]

    PoAct: Policy and Action Dual-Control Agent for Generalized Applications

    Guozhi Yuan, Youfeng Liu, Jingli Yang, Wei Jia, Kai Lin, Yansong Gao, Shan He, Zilin Ding, and Haitao Li. Poact: Policy and action dual-control agent for generalized applications.arXiv preprint arXiv:2501.07054, 2025

    work page Pith review arXiv 2025

  6. [6]

    Compass: Enhancing agent long-horizon reasoning with evolving context.arXiv preprint arXiv:2510.08790, 2025

    Guangya Wan, Mingyang Ling, Xiaoqi Ren, Rujun Han, Sheng Li, and Zizhao Zhang. Compass: Enhancing agent long-horizon reasoning with evolving context.arXiv preprint arXiv:2510.08790, 2025

    arXiv 2025

  7. [7]

    Pre-act: Multi-step planning and reasoning improves acting in llm agents.arXiv preprint arXiv:2505.09970, 2025

    Mrinal Rawat, Ambuje Gupta, Rushil Goomer, Alessandro Di Bari, Neha Gupta, and Roberto Pieraccini. Pre-act: Multi-step planning and reasoning improves acting in llm agents.arXiv preprint arXiv:2505.09970, 2025

    arXiv 2025

  8. [8]

    Hipporag: Neurobiologically inspired long-term memory for large language models.Advances in neural information processing systems, 37:59532–59569, 2024

    Bernal J Gutiérrez, Yiheng Shu, Yu Gu, Michihiro Yasunaga, and Yu Su. Hipporag: Neurobiologically inspired long-term memory for large language models.Advances in neural information processing systems, 37:59532–59569, 2024

    2024

Show all 74 references
  1. [9]

    Transformer-xl: Attentive language models beyond a fixed-length context

    Zihang Dai, Zhilin Yang, Yiming Yang, Jaime G Carbonell, Quoc Le, and Ruslan Salakhutdinov. Transformer-xl: Attentive language models beyond a fixed-length context. InProceedings of the 57th annual meeting of the association for computational linguistics, pages 2978–2988, 2019

    2019

  2. [10]

    Abbel: Llm agents acting through belief bottlenecks expressed in language.arXiv preprint arXiv:2512.20111, 2025

    Aly Lidayan, Jakob Bjorner, Satvik Golechha, Kartik Goyal, and Alane Suhr. Abbel: Llm agents acting through belief bottlenecks expressed in language.arXiv preprint arXiv:2512.20111, 2025

    2025 arXiv

  3. [11]

    Ai planning framework for llm-based web agents.arXiv preprint arXiv:2603.12710, 2026

    Orit Shahnovsky and Rotem Dror. Ai planning framework for llm-based web agents.arXiv preprint arXiv:2603.12710, 2026

    2026

  4. [12]

    Do llms recognize your preferences? evaluating personalized preference following in llms.arXiv preprint arXiv:2502.09597, 2025

    Siyan Zhao, Mingyi Hong, Yang Liu, Devamanyu Hazarika, and Kaixiang Lin. Do llms recognize your preferences? evaluating personalized preference following in llms.arXiv preprint arXiv:2502.09597, 2025

    2025 arXiv

  5. [13]

    Doing personal laps: Llm-augmented dialogue construction for personalized multi-session conversational search

    Hideaki Joko, Shubham Chatterjee, Andrew Ramsay, Arjen P De Vries, Jeff Dalton, and Faegheh Hasibi. Doing personal laps: Llm-augmented dialogue construction for personalized multi-session conversational search. InProceedings of the 47th International ACM SIGIR Conference on Re...

    2024

  6. [14]

    Towards proactive personalization through profile customization for individual users in dialogues.arXiv preprint arXiv:2512.15302, 2025

    Xiaotian Zhang, Yuan Wang, Ruizhe Chen, Zeya Wang, Runchen Hou, and Zuozhu Liu. Towards proactive personalization through profile customization for individual users in dialogues.arXiv preprint arXiv:2512.15302, 2025

    2025

  7. [15]

    Towards realistic personalization: Evaluating long- horizon preference following in personalized user-llm interactions.arXiv preprint arXiv:2603.04191, 2026

    Qianyun Guo, Yibo Li, Yue Liu, and Bryan Hooi. Towards realistic personalization: Evaluating long- horizon preference following in personalized user-llm interactions.arXiv preprint arXiv:2603.04191, 2026

    2026

  8. [16]

    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 (Volume 1: Long Papers), p...

    2024

  9. [17]

    Longmemeval: Benchmarking chat assistants on long-term interactive memory.arXiv preprint arXiv:2410.10813, 2024

    Di Wu, Hongwei Wang, Wenhao Yu, Yuwei Zhang, Kai-Wei Chang, and Dong Yu. Longmemeval: Benchmarking chat assistants on long-term interactive memory.arXiv preprint arXiv:2410.10813, 2024

    2024 arXiv

  10. [18]

    Augmenting language models with long-term memory.Advances in Neural Information Processing Systems, 36: 74530–74543, 2023

    Weizhi Wang, Li Dong, Hao Cheng, Xiaodong Liu, Xifeng Yan, Jianfeng Gao, and Furu Wei. Augmenting language models with long-term memory.Advances in Neural Information Processing Systems, 36: 74530–74543, 2023. 11

    2023

  11. [19]

    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. InProceedings of the AAAI conference on artificial intelligence, volume 38, pages 19724–19731, 2024

    2024

  12. [20]

    Memory os of ai agent

    Jiazheng Kang, Mingming Ji, Zhe Zhao, and Ting Bai. Memory os of ai agent. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 25972–25981, 2025

    2025

  13. [21]

    Mem0: Building production-ready ai agents with scalable long-term memory.arXiv preprint arXiv:2504.19413, 2025

    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

    2025 arXiv

  14. [22]

    A-mem: Agentic memory for llm agents.arXiv preprint arXiv:2502.12110, 2025

    Wujiang Xu, Zujie Liang, Kai Mei, Hang Gao, Juntao Tan, and Yongfeng Zhang. A-mem: Agentic memory for llm agents.arXiv preprint arXiv:2502.12110, 2025

    2025 arXiv

  15. [23]

    Generative agents: Interactive simulacra of human behavior

    Joon Sung Park, Joseph O’Brien, Carrie Jun Cai, Meredith Ringel Morris, Percy Liang, and Michael S Bernstein. Generative agents: Interactive simulacra of human behavior. InProceedings of the 36th annual acm symposium on user interface software and technology, pages 1–22, 2023

    2023

  16. [24]

    Retrieval-augmented generation for knowledge-intensive nlp tasks.Advances in neural information processing systems, 33:9459–9474, 2020

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, et al. Retrieval-augmented generation for knowledge-intensive nlp tasks.Advances in neural information processing systems,...

    2020

  17. [25]

    Medmt-bench: Can llms memo- rize and understand long multi-turn conversations in medical scenarios?arXiv preprint arXiv:2603.23519, 2026

    Lin Yang, Yuancheng Yang, Xu Wang, Changkun Liu, and Haihua Yang. Medmt-bench: Can llms memo- rize and understand long multi-turn conversations in medical scenarios?arXiv preprint arXiv:2603.23519, 2026

    2026

  18. [26]

    Patil, Ion Stoica, and Joseph E

    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

    2023 arXiv

  19. [27]

    From isolated conversations to hierarchical schemas: Dynamic tree memory representation for llms.arXiv preprint arXiv:2410.14052, 2024

    Alireza Rezazadeh, Zichao Li, Wei Wei, and Yujia Bao. From isolated conversations to hierarchical schemas: Dynamic tree memory representation for llms.arXiv preprint arXiv:2410.14052, 2024

    2024 arXiv

  20. [28]

    Zep: a temporal knowledge graph architecture for agent memory.arXiv preprint arXiv:2501.13956, 2025

    Preston Rasmussen, Pavlo Paliychuk, Travis Beauvais, Jack Ryan, and Daniel Chalef. Zep: a temporal knowledge graph architecture for agent memory.arXiv preprint arXiv:2501.13956, 2025

    2025 arXiv

  21. [29]

    Towards lifelong dialogue agents via timeline-based memory management

    Kai Tzu-iunn Ong, Namyoung Kim, Minju Gwak, Hyungjoo Chae, Taeyoon Kwon, Yohan Jo, Seung-won Hwang, Dongha Lee, and Jinyoung Yeo. Towards lifelong dialogue agents via timeline-based memory management. InProceedings of the 2025 Conference of the Nations of the Americas Chapter ...

    2025

  22. [30]

    In prospect and retrospect: Reflective memory management for long-term per- sonalized dialogue agents

    Zhen Tan, Jun Yan, I-Hung Hsu, Rujun Han, Zifeng Wang, Long Le, Yiwen Song, Yanfei Chen, Hamid Palangi, George Lee, et al. In prospect and retrospect: Reflective memory management for long-term per- sonalized dialogue agents. InProceedings of the 63rd Annual Meeting of the Ass...

    2025

  23. [31]

    Bridging intuitive associations and deliberate recall: Empowering llm personal assistant with graph-structured long-term memory

    Yujie Zhang, Weikang Yuan, and Zhuoren Jiang. Bridging intuitive associations and deliberate recall: Empowering llm personal assistant with graph-structured long-term memory. InFindings of the Association for Computational Linguistics: ACL 2025, pages 17533–17547, 2025

    2025

  24. [32]

    Enhancing retrieval-augmented large language models with iterative retrieval-generation synergy

    Zhihong Shao, Yeyun Gong, Yelong Shen, Minlie Huang, Nan Duan, and Weizhu Chen. Enhancing retrieval-augmented large language models with iterative retrieval-generation synergy. InFindings of the Association for Computational Linguistics: EMNLP 2023, pages 9248–9274, 2023

    2023

  25. [33]

    Raptor: Recursive abstractive processing for tree-organized retrieval

    Parth Sarthi, Salman Abdullah, Aditi Tuli, Shubh Khanna, Anna Goldie, and Christopher D Manning. Raptor: Recursive abstractive processing for tree-organized retrieval. InThe Twelfth International Conference on Learning Representations, 2024

    2024

  26. [34]

    From local to global: A graph rag approach to query-focused summarization.arXiv preprint arXiv:2404.16130, 2024

    Darren Edge, Ha Trinh, Newman Cheng, Joshua Bradley, Alex Chao, Apurva Mody, Steven Truitt, Dasha Metropolitansky, Robert Osazuwa Ness, and Jonathan Larson. From local to global: A graph rag approach to query-focused summarization.arXiv preprint arXiv:2404.16130, 2024

    2024 arXiv

  27. [35]

    LangMem: Long-term memory for ai agents

    LangChain Team. LangMem: Long-term memory for ai agents. https://langchain-ai.github.io/ langmem/, 2025. Software documentation. Accessed: 2026-05-06

    2025

  28. [36]

    Bleu: 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

    2002

  29. [37]

    LGBTQ support group

    Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in neural information processing systems, 36:46595–46623, 2023. 12 APPENDIX...

    2023

  30. [38]

    Use only the supplied claims, facets, exact terms, and recent snapshot notes

  31. [39]

    Preserve exact titles, books, recipes, instruments, symbols, places, and named items

  32. [40]

    Keep list items explicit; do not collapse them into broad themes

  33. [41]

    Do not let the newest update erase older item sets, places, events, dates, or counts from the trajectory

  34. [42]

    sunset" visible even if you also mention

    Do not replace concrete source phrases with broad categories: keep "sunset" visible even if you also mention "artwork"

  35. [43]

    Mention contradictory or uncertain facts under Conflicts / Uncertainty instead of resolving them yourself

  36. [44]

    Do not add commentary, preamble, or conclusions outside the markdown sections

  37. [45]

    Wiki Page Planning

    Exclude filler words, acknowledgements, discourse markers, and casual fragments from the useful retrieval content. Wiki Page Planning. 25 TASK=WIKI_PAGE_PLAN Plan a compact sample-level wiki graph from the provided candidate seed manifest. Return markdown only with the exact h...

  38. [46]

    Always include exactly one index page

  39. [47]

    Treat each input block as a candidate seed that already groups related trajectories; refine or select from those seeds instead of inventing broad new groupings

  40. [48]

    Create entity pages when a person/entity appears across multiple trajectories

  41. [49]

    Create inventory pages for list/count/item-heavy facts such as activities, books, recipes, instruments, symbols, places, wins, dogs, or repeated events

  42. [50]

    Create topic pages only when they add navigation value beyond entity or inventory pages

  43. [51]

    Keep pages compact and specific; avoid broad pages that overlap heavily with existing entity or inventory coverage

  44. [52]

    Prefer explicit evidence coverage over narrative grouping

  45. [53]

    Use trajectory historical evidence cards as first-class evidence; do not rely only on a trajectory's latest summary

  46. [54]

    Wiki Page Compilation

    If a card exposes a specific item, place, event, count, or source anchor, cover it in a non-index page whenever possible. Wiki Page Compilation. TASK=WIKI_PAGE_COMPILE Compile one wiki page from the supplied page metadata, representative trajectory summaries, and historical tr...

  47. [55]

    Use only supplied linked trajectory ids, dominant entities, exact terms, facet values, representative trajectory summaries, and historical evidence cards

  48. [56]

    Keep exact names, items, places, counts, and titles

  49. [57]

    For inventory pages, favor explicit item lists over prose

  50. [58]

    Preserve enumerated items and counts explicitly; do not collapse them into broad narrative summaries

  51. [59]

    sunset", keep

    Do not replace a concrete source phrase with a broader category: if an evidence card contains "sunset", keep "sunset", not only "nature-inspired artwork"

  52. [60]

    Do not add speculative links or inferred facts not stated in the supplied inputs

  53. [61]

    In Items / Counts, use only readable display items, readable counts, and readable key facts

  54. [62]

    If a latest summary and historical card emphasize different facts, present the historical facts in their own concrete bullets instead of dropping them

  55. [63]

    Do not dump Dominant exact terms verbatim; ignore any low-confidence or fragment-like signal

  56. [64]

    Not provided

    Do not write placeholder descriptions such as "Not provided", "Unknown", "N/A", "None provided", "No specific key facts", or "No explicit items". 26

  57. [65]

    Describe a linked trajectory only when its representative summary or evidence card is supplied; otherwise leave the trajectory id without a fabricated description

  58. [66]

    C.3 Answer Generation and Evaluation Prompts LoCoMo Evidence Synthesis

    The system may rewrite the final Linked Trajectories section deterministically, so keep that section concise. C.3 Answer Generation and Evaluation Prompts LoCoMo Evidence Synthesis. TASK=LoCoMo_ANSWER_EVIDENCE_SYNTHESIS Legacy structured LoCoMo evidence synthesis. Input contai...

  59. [67]

    If it cannot answer the question, set can_answer= false, leave final_answer empty, and explain the gap in abstain_reason

    Use only retrieved evidence. If it cannot answer the question, set can_answer= false, leave final_answer empty, and explain the gap in abstain_reason

  60. [68]

    Preserve exact source terms for names, titles, places, dates, counts, activities, and items

  61. [69]

    supporting_source_refs must use only visible refs such as D8:4, and each ref must support the requested answer family

  62. [70]

    Exclude future plans, intentions, reactions, duplicate mentions, general hobbies, and uncertain candidates

    Count distinct completed events only. Exclude future plans, intentions, reactions, duplicate mentions, general hobbies, and uncertain candidates

  63. [71]

    Count final_answer must be natural language, not a bare number; use an exact count only when complete evidence supports it, otherwise state a retrieved- evidence lower bound

  64. [72]

    yesterday

    Date/time answers must use source line date=... fields, snapshot Timestamp lines, and the ## Temporal Anchors block; resolve terms such as "yesterday" to the grounded date when available

  65. [73]

    For lists/events/preferences, include every in-scope source-backed specific item and avoid scope-mismatched extras

  66. [74]

    West County

    For bridge/alias facts, answer with the concrete source-backed value, e.g. " West County" rather than "old area". Return structured data only. If structured output is unavailable, return only a JSON object matching the same fields. LoCoMo Judge. Evaluate the candidate answer a...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.