Pith. sign in

REVIEW 5 major objections 4 minor 26 references

A unified two-layer memory store—narrative profiles plus co-extracted exact facts—lets LLM agents traverse multi-hop entity chains without building a knowledge graph.

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

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

T0 review · deepseek-v4-flash

2026-08-02 12:32 UTC pith:SWQSPFVD

load-bearing objection A useful, honest memory-architecture paper whose central specialization claim needs one more round of ablation and an alias-robustness check. the 5 major comments →

arxiv 2607.19359 v1 pith:SWQSPFVD submitted 2026-06-01 cs.AI

Profile-Graph Memory for LLM Agents: Implicit Cross-Entity Traversal through Narrative Profiles

Shengtong Zhu This is my paper
classification cs.AI
keywords LLM agentslong-term memorymulti-hop reasoningmemory benchmarkprofile summarizationretrieval augmentationimplicit knowledge graphcompression residuals
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper sets out to show that long-term memory for LLM agents can handle both multi-hop association and precision recall in one store, without an explicit knowledge graph built at write time. The proposed ProGraph architecture keeps a narrative profile per entity and, in the same LLM write call, a set of compression residuals—exact dates, quantities, and named items that the summary would otherwise paraphrase away. At read time it retrieves profiles by embedding, then expands to neighboring entities by scanning profile text for registered entity names, treating co-mentions as implicit edges; residuals of selected entities are ranked and appended as verified facts. On MemHop, a new 1,000-question benchmark spanning hop depths 1–5, ProGraph reaches 80.1%, matching the full-conversation baseline; on LoCoMo it reaches 78.4%, 11.3 points above that baseline. The ablation claims a clean division of labor: profile expansion drives multi-hop accuracy (−22.6 points when removed on MemHop), while residual co-extraction drives precision (−8.6 points when removed on LoCoMo), with cross-effects under 3 points.

Core claim

The central claim is mechanism specialization. ProGraph writes two layers per entity in a single LLM call: a narrative profile (for association) and compression residuals (for precision). Retrieval is three embedding-only stages—profile retrieval, relevance-gated expansion that scans retrieved profiles for registered entity names and pulls qualifying neighbors into the answer context (iterated up to five times), and residual-augmented context assembly. The paper's evidence for specialization is a full-grid ablation: disabling expansion costs 22.6 points on MemHop but only 0.7 on LoCoMo; disabling write-time residual extraction costs 8.6 points on LoCoMo but only 2.9 on MemHop. End-to-end, th

What carries the argument

The load-bearing mechanism is the implicit entity graph formed by substring co-mention inside LLM-written profile narratives: when an LLM writes Alice's profile as 'Alice lives with her roommate Bob…', the registered string 'Bob' appears inside Alice's profile, creating an edge Alice→Bob with no relation-extraction step. Stage-2 relevance-gated expansion operationalizes this by scanning selected profiles for registered entity names (full-name substring or first-token word match), scoring each candidate neighbor's profile against the query embedding, adding it only if cosine similarity clears a fixed gate, and iterating up to five times. The complement is compression residuals—short atomic fa

Load-bearing premise

The central result depends on profile narratives always naming bridge-step neighbors by their registered full name or first token, so that expansion fires by substring matching alone; if a profile refers to the neighbor only as 'my roommate' or by an alias, the implicit edge is invisible to the store.

What would settle it

Take the released MemHop data and check every K≥2 gold chain: does the predecessor profile contain the bridge entity's registered name as a substring or first-token whole word? Any chain that fails is unanswerable by ProGraph's expansion by construction. Then run the expansion on/off ablation on the subset where all bridges satisfy the naming assumption; if the 22.6-point gain largely disappears, string-matched traversal was not the operative mechanism.

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

If this is right

  • One memory store can serve both multi-hop associative queries and precision-sensitive single-hop queries without per-query routing; the two mechanisms are decoupled by benchmark, with cross-effects under 3 points.
  • Explicit knowledge-graph construction is not required for conversational multi-hop memory: a language-level traversal over narrative co-mentions outperforms the graph-based baseline by large margins on both benchmarks.
  • Stage-2 expansion acts partly as a retrieval rescue, pulling answer-bearing profiles that cosine ranked below top-M into context; this explains why expansion helps even K=1 questions and why the gain shrinks and even reverses at K=5 under a larger expansion cap.
  • Short, entity-bound retrieval units improve refusal behavior on unanswerable (adversarial) questions, rather than tempting the answer LLM to hallucinate from peripheral context.
  • Failure analysis locates most remaining MemHop errors in multi-hop traversal (wrong bridge entity) rather than precision, and refusals concentrate at K=1–2, suggesting an answer-LLM mode-switching effect.

Where Pith is reading between the lines

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

  • If mechanism specialization holds beyond these two benchmarks, memory budgets could be tuned per layer: spend write-time tokens on residual extraction when precision queries dominate, and on expansion depth when association queries dominate, without altering the retrieval skeleton.
  • The string-match assumption is the most fragile point for transfer to real conversations: people are often referred to by role or alias. Relaxing expansion to resolve role phrases ('my roommate') or aliases before substring matching would test whether the −22.6-point expansion gain survives noisier naming.
  • MemHop's synthetic construction nearly guarantees registered names appear in profiles; a natural variant with elliptical or paraphrased bridge references would measure how much of the expansion contribution is an artifact of clean naming.
  • The observed K=5 regression under Hmax=5 suggests an adaptive expansion cap (for example, gated by query specificity or fan-out size) could recover deep-chain accuracy; the paper leaves this as future work but its own sensitivity table shows the crossover point.

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 / 4 minor

Summary. The paper proposes Profile-Graph Memory (ProGraph), a two-layer LLM-agent memory architecture consisting of entity-centric narrative profiles (for association) and 'compression residuals' (precision-critical atomic facts) that are co-extracted with each profile update in a single LLM call. Retrieval is pure-embedding: profile retrieval with a name boost, relevance-gated Stage-2 expansion that scans profiles for exact registered entity-name substrings and pulls matched neighbor profiles/residuals into the context, and residual augmentation. The paper also introduces MemHop, a synthetic 1,000-question multi-hop conversational memory benchmark with per-hop decompositions and gold evidence. Experiments report that ProGraph matches FullContext on MemHop (80.1 vs. 79.6) and exceeds it on LoCoMo (78.4 vs. 67.1 LightMem; 71.3 vs. 61.0 strict), while outperforming Mem0, A-Mem, HippoRAG, and RAG. An ablation claims mechanism specialization: expansion carries MemHop (-22.6pp) and write-time residuals carry LoCoMo (-8.6pp), with cross-effects under 3pp. The paper releases MemHop, ProGraph, baseline implementations, and construction code, and includes a candid Limitations section and error analysis.

Significance. If the mechanism-specialization result holds, the paper makes a useful contribution: MemHop fills a real gap in multi-hop conversational memory evaluation, and ProGraph offers a low-cost way to obtain multi-hop traversal without explicit KG construction and precision recall without a second extraction call. Strengths include the released benchmark with per-hop evidence, the automated QA pipeline, the zero-extra-API-call co-extraction design, and the unusually detailed Limitations and error decomposition. However, the central architecture-level claim is not yet fully established: the ablation table is not actually a full grid, the expansion mechanism's dependence on exact entity-name substrings is unmeasured, and the Hmax sensitivity data suggest much of the expansion gain is a Stage-1 'rescue' effect rather than cross-entity traversal. These issues are addressable with additional experiments, but they are load-bearing for the claimed mechanism specialization.

major comments (5)
  1. [§5.2, Table 5] The abstract and §5.2 call Table 5 a 'full-grid ablation,' but only five of the 2^3 = 8 switch combinations (Exp × R-time × W-time) are shown. The missing cells include the all-off cell and the expansion-off/write-time-off interaction, so the 'cross-effects ≤3pp' and 'decoupled mechanisms' conclusions are not supported by the displayed data. In addition, the LoCoMo column of Table 5 uses the strict judge (full system 71.3, matching Table 3), while the main LoCoMo results and the abstract use LightMem (78.4); the -8.6pp write-time-residual effect is therefore only demonstrated under the secondary judge. Please run the complete 2^3 grid and report both judges.
  2. [§3.2, Appendix A] Stage-2 expansion fires only on exact full-name substring or first-token whole-word matches. Appendix A asserts that profiles 'almost always' surface entity names in surface form, but the paper provides no measurement of how often bridge entities are actually named in MemHop profiles, and the Limitations do not discuss aliases or role references ('my roommate', 'the chef', etc.). Because MemHop's dialogues and profiles are machine-generated with consistent name usage, the -22.6pp expansion ablation may overstate the mechanism's external validity for real conversational narratives. Please report exact-name mention coverage across MemHop's evidence chains, and add a targeted alias/role-reference robustness test (or explicitly bound the scope of the claim).
  3. [Appendix D] The Hmax sensitivity analysis substantially complicates the mechanism-specialization story. Raising Hmax from 1 to 5 gives +12.8pp at K=1 and a negative change at K=5, and the authors attribute this to a Stage-1 rescue effect rather than chain traversal. The -22.6pp expansion-off ablation therefore conflates rescue (answer-bearing profile pulled in via a neighbor's mention) with genuine cross-entity bridge traversal. Please report per-K expansion on/off deltas and, if possible, decompose the expansion contribution into rescue vs. traversal by labeling whether the answer entity was already in the Stage-1 selection set.
  4. [§5, Hyperparameters] The architecture has at least six hand-tuned parameters (τ_dedup, τ_expand, top-M, Hmax, β, residual cap), tuned on held-out data, but the paper gives no split definition, selection protocol, or sensitivity analysis except for Hmax. The ablation deltas are threshold-dependent: a different τ_expand or top-M will change how much rescue vs. fan-out contributes, and hence the measured expansion effect. Please provide the tuning split and one-factor-at-a-time sensitivity for the key thresholds, or otherwise justify robustness of the ablation conclusions to these choices.
  5. [§5.2, Table 5] The paper states that 95% bootstrap CIs are ±2–3pp for n≥1,000 but reports no intervals for the ablation cells. Under that uncertainty, the -0.9, -2.9, and -0.7pp effects are within noise; only -22.6 and -8.6 are clearly nonzero. Since the 'cross-effects under 3pp' is a central quantitative claim, report CIs for Table 5 (ideally with multiple seeds or conversation splits) before claiming that the mechanisms are decoupled.
minor comments (4)
  1. [Table 5] The row labels are ambiguous in the rendered table. Use explicit on/off markers (✓/✗) for each of Exp, R-time, and W-time in every row so the reader can identify which condition is compared against which.
  2. [§5.2] The 'Methodological note' discusses the read-time switch, but the corresponding row in Table 5 is not clearly identified. Refer to specific row conditions when explaining the -0.9pp read-time result.
  3. [Limitations] The sentence 'a pilot run confirms our pipeline produces sensible LongMemEval predictions' is unverifiable as written. Either report the pilot numbers in an appendix or remove the sentence.
  4. [Table 3 caption] The discussion of F1 and answer verbosity is helpful, but the token-F1 comparison is harsh on systems with longer answers. Consider also reporting length-controlled metrics or citing a standard reference for this known issue.

Circularity Check

1 steps flagged

MemHop's multi-hop questions are defined as entity-name bridge chains, and Stage 2 expansion is defined as substring-matched entity-name traversal, so the −22.6pp expansion ablation is partly a by-construction result; the LoCoMo precision side is external and independent.

specific steps
  1. self definitional [§1 (multi-hop as entity-chain traversal); §3.2 + Appendix A (expansion by entity-name substring match); §4 'Entity bridge chain method'; §5.2 ablation]
    "Multi-hop association: answering “what instrument does Alice's roommate play?” requires traversing an entity chain Alice → Bob → piano ... We generate K-hop questions by sampling a chain of K+1 entities linked by K relationships, e1 →e 2 → · · · →eK+1 ... we scan profiles in S for already-registered entity names (Appendix A) ... An entity name n counts as mentioned in profile pi if (i) the full name appears as a case-insensitive substring or (ii) its first token appears as a whole word ... On MemHop, disabling expansion costs 22.6pp."

    MemHop operationalizes 'multi-hop' by construction as a chain of named entities: a question starts at e1 and asks for an attribute of eK+1, with each hop resolved through the named entities in the chain. ProGraph's Stage 2 expansion is defined as the same operation: scanning narrative profiles for registered entity-name substrings/whole words and adding those entities to the retrieval set. Therefore the −22.6pp ablation measures the importance of the entity-name co-mention matcher on a benchmark whose questions were generated to be answerable by exactly that matcher. The 'expansion drives multi-hop reasoning' claim is thus partly a definitional tautology: the benchmark's target operation and the ablated component are the same operation. The LoCoMo expansion effect (−0.7pp) and the residual

full rationale

The paper contains no equation-level circularity: ProGraph's outputs are not fitted to the reported scores, and the LoCoMo results are measured against an external benchmark with standard baselines. The residual/precision side of the central 'mechanism specialization' claim is therefore independently supported. The serious concern is confined to the MemHop side. MemHop is introduced in the same paper and constructed by sampling entity-name bridge chains; ProGraph's headline associative mechanism, Stage 2 expansion, is precisely substring/whole-token matching of registered entity names in profile text. The −22.6pp expansion ablation then reports the contribution of 'entity-name chain traversal' on a benchmark whose questions are, by construction, entity-name chains. This is not a statistical artifact but a construct-level circularity: the benchmark's notion of multi-hop is defined as the very operation the ablated component performs. Appendix A's assumption that profiles 'almost always surface entity names in surface form' is asserted, not measured, and no alias/role-reference robustness test is reported, so the transferability of the MemHop result to natural narratives is unestablished. These are external-validity and construct-validity issues rather than hidden fitting, but they reduce the force of the strongest claim. Weighted against the independent LoCoMo evidence, I assign a partial-circularity score of 6.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 0 invented entities

No new physical or formal entities are postulated; ProGraph and MemHop are artifacts, not entities requiring independent evidence. Free parameters are the hand-tuned retrieval hyperparameters that the paper itself flags as deployment-sensitive.

free parameters (6)
  • tau_dedup = 0.9
    Residual dedup threshold; stated in §3.1, hand-tuned.
  • tau_expand = 0.2
    Relevance gate for Stage-2 expansion; stated in §3.2; controls fan-out; hand-tuned.
  • top_M = 5
    Number of profiles selected in Stage 1; stated in §3.2; hand-tuned.
  • Hmax = 5
    Stage-2 expansion hop cap; sweep only over {1,5}; stated §3.2/Appendix D.
  • name_boost_beta = 0.3
    Boost applied to entities explicitly named in query; §3.2; hand-tuned.
  • residual_cap = not specified
    Per-entity cap on residuals appended at Stage 3; §3.2; value not reported.
axioms (4)
  • domain assumption Entity names in profiles are consistently surface-form (exact substring or first token matchable via registered name registry).
    Load-bearing for Stage-2 expansion; Appendix A defines matching but does not test robustness to alias/paraphrase variation.
  • domain assumption GPT-4o-mini serves as both answer LLM and judge with symmetric bias across systems; LightMem judge is a valid correctness proxy.
    Evaluation depends on this; no human eval or inter-annotator agreement reported.
  • domain assumption MemHop's synthetic dialogues faithfully realize the underlying relationship graph, and its 1,000 questions are representative of multi-hop conversational memory.
    Benchmark is authored by the same paper; construction prompts released but no independent validation of realism.
  • ad hoc to paper Co-extracting residuals in the same call does not reduce profile quality; the -8.6pp write-time ablation attributes the entire loss to missing residuals rather than to a degraded profile.
    The ablation design assumes the joint prompt is neutral; profile-only prompt may produce different narratives, confounding the switch.

pith-pipeline@v1.3.0-alltime-deepseek · 13171 in / 13453 out tokens · 131612 ms · 2026-08-02T12:32:38.492002+00:00 · methodology

0 comments
read the original abstract

Long-term memory is essential for LLM agents that interact across sessions, yet current memory benchmarks primarily evaluate single-hop recall, leaving multi-hop association largely unmeasured. We make three contributions. First, we introduce MemHop, a multi-hop memory benchmark of 1,000 questions at hop depths 1-5 across 10 social-network scenarios, with per-hop evidence annotations. Second, we present Profile-Graph Memory (ProGraph), a two-layer memory architecture combining (i) profile expansion -- substring-matched traversal of entity names that naturally appear in LLM-written profile narratives, a minimal alternative to explicit knowledge-graph construction -- and (ii) compression residuals -- exact dates, quantities, and named items co-extracted with each profile update at zero extra API cost. Third, a full-grid ablation shows cross-benchmark mechanism specialization: profile expansion drives multi-hop reasoning (-22.6pp on MemHop when removed) while compression residuals drive precision recall (-8.6pp on LoCoMo when not co-extracted), with cross-effects under 3pp within a single architecture. ProGraph averages 80.1% on MemHop (matching the FullContext reference) and 78.4% on LoCoMo (exceeding FullContext by 11.3pp), outperforming Mem0, A-Mem, HippoRAG, and RAG on both. We release MemHop, ProGraph, and baseline implementations.

Figures

Figures reproduced from arXiv: 2607.19359 by Shengtong Zhu.

Figure 1
Figure 1. Figure 1: Profile-Graph Memory architecture. Write (left): each session is parsed for entities, and for each entity a single LLM call produces the updated profile and the residuals it could not preserve, at zero extra API cost. Read (right): three pure-embedding stages (profile retrieval → relevance-gated expansion over entity names in profile text → residual-augmented context assembly) feed one answer-generation ca… view at source ↗
Figure 2
Figure 2. Figure 2: Accuracy (%) vs. hop depth K on MemHop (10-scenario average, n=1,000). in store) makes residuals look near-inert (−0.9pp on both benchmarks) because the joint extraction prompt lets profiles absorb the same facts. Only the write-time switch (profile-only extraction prompt) reveals what residuals contribute when the profile must stand alone—an order of magnitude larger on LoCoMo. 5.3 Accuracy vs. hop depth … view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

26 extracted references · 7 linked inside Pith

  1. [1]

    Luiz C Borro, Luiz AB Macarini, Gordon Tindall, Michael Montero, and Adam B Struck. 2026. Memori: A persistent memory layer for efficient, context-aware llm agents. arXiv preprint arXiv:2603.19935

    arXiv 2026

  2. [2]

    Prateek Chhikara, Dev Khant, Saket Aryan, Taranjeet Singh, and Deshraj Yadav. 2025. Mem0: Building production-ready ai agents with scalable long-term memory. arXiv preprint arXiv:2504.19413

    Pith/arXiv arXiv 2025

  3. [3]

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

    Pith/arXiv arXiv 2024

  4. [4]

    Jizhan Fang, Xinle Deng, Haoming Xu, Ziyan Jiang, Yuqi Tang, Ziwen Xu, Shumin Deng, Yunzhi Yao, Mengru Wang, Shuofei Qiao, and 1 others. 2025. Lightmem: Lightweight and efficient memory-augmented generation

    2025

  5. [5]

    Zirui Guo, Lianghao Xia, Yanhua Yu, Tian Ao, and Chao Huang. 2024. Lightrag: Simple and fast retrieval-augmented generation. arXiv preprint arXiv:2410.05779, 2(3)

    Pith/arXiv arXiv 2024

  6. [6]

    Bernal J Guti \'e rrez, Yiheng Shu, Yu Gu, Michihiro Yasunaga, and Yu Su. 2024. Hipporag: Neurobiologically inspired long-term memory for large language models. volume 37, pages 59532--59569

    2024

  7. [7]

    Xanh Ho, Anh-Khoa Duong Nguyen, Saku Sugawara, and Akiko Aizawa. 2020. Constructing a multi-hop qa dataset for comprehensive evaluation of reasoning steps. In Proceedings of the 28th International Conference on Computational Linguistics, pages 6609--6625

    2020

  8. [8]

    Chuanrui Hu, Xingze Gao, Zuyi Zhou, Dannong Xu, Yi Bai, Xintong Li, Hui Zhang, Tong Li, Chong Zhang, Lidong Bing, and 1 others. 2026. Evermemos: A self-organizing memory operating system for structured long-horizon reasoning. arXiv preprint arXiv:2601.02163

    arXiv 2026

  9. [9]

    Chris Latimer, Nicol \'o Boschi, Andrew Neeser, Chris Bartholomew, Gaurav Srivastava, Xuan Wang, and Naren Ramakrishnan. 2025. Hindsight is 20/20: Building agent memory that retains, recalls, and reflects. arXiv preprint arXiv:2512.12818

    arXiv 2025

  10. [10]

    Jiaqi Liu, Yaofeng Su, Peng Xia, Siwei Han, Zeyu Zheng, Cihang Xie, Mingyu Ding, and Huaxiu Yao. 2026. Simplemem: Efficient lifelong memory for llm agents. arXiv preprint arXiv:2601.02553

    Pith/arXiv arXiv 2026

  11. [11]

    Adyasha Maharana, Dong-Ho Lee, Sergey Tulyakov, Mohit Bansal, Francesco Barbieri, and Yuwei Fang. 2024. Evaluating very long-term conversational memory of llm agents. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 13851--13870

    2024

  12. [12]

    OpenAI . 2024. https://openai.com/index/gpt-4o-mini-advancing-cost-efficient-intelligence/ Gpt-4o mini

    2024

  13. [13]

    Charles Packer, Vivian Fang, Shishir\_G Patil, Kevin Lin, Sarah Wooders, and Joseph\_E Gonzalez. 2023. Memgpt: towards llms as operating systems. ArXiv

    2023

  14. [14]

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

    2023

  15. [15]

    Nils Reimers and Iryna Gurevych. 2019. Sentence-bert: Sentence embeddings using siamese bert-networks. In Proceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conference on natural language processing (EMNLP-IJCNLP), pages 3982--3992

    2019

  16. [16]

    Miao Su, Yucan Guo, Zhongni Hou, Long Bai, Zixuan Li, Yufei Zhang, Guojun Yin, Wei Lin, Xiaolong Jin, Jiafeng Guo, and 1 others. 2026. Beyond dialogue time: Temporal semantic memory for personalized llm agents. arXiv preprint arXiv:2601.07468

    arXiv 2026

  17. [17]

    Jingwei Sun, Jianing Zhu, Jiangchao Yao, Tongliang Liu, and Bo Han. 2026. Rethinking how to remember: Beyond atomic facts in lifelong llm agent memory. arXiv preprint arXiv:2605.19952

    Pith/arXiv arXiv 2026

  18. [18]

    Harsh Trivedi, Niranjan Balasubramanian, Tushar Khot, and Ashish Sabharwal. 2022. MuSiQue : Multihop questions via single-hop question composition. volume 10, pages 539--554. MIT Press

    2022

  19. [19]

    Lei Wei, Xu Dong, Xiao Peng, Niantao Xie, and Bin Wang. 2026. Fademem: Biologically-inspired forgetting for efficient agent memory. arXiv preprint arXiv:2601.18642

    arXiv 2026

  20. [20]

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

    Pith/arXiv arXiv 2024

  21. [21]

    Wujiang Xu, Zujie Liang, Kai Mei, Hang Gao, Juntao Tan, and Yongfeng Zhang. 2026. A-mem: Agentic memory for llm agents. volume 38, pages 17577--17604

    2026

  22. [22]

    Shannan Yan, Jingchen Ni, Leqi Zheng, Jiajun Zhang, Peixi Wu, Dacheng Yin, Jing Lyu, Chun Yuan, and Fengyun Rao. 2026. Adamem: Adaptive user-centric memory for long-horizon dialogue agents. arXiv preprint arXiv:2603.16496

    Pith/arXiv arXiv 2026

  23. [23]

    Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William Cohen, Ruslan Salakhutdinov, and Christopher D Manning. 2018. Hotpotqa: A dataset for diverse, explainable multi-hop question answering. In Proceedings of the 2018 conference on empirical methods in natural language processing, pages 2369--2380

    2018

  24. [24]

    Guilin Zhang, Wei Jiang, Xiejiashan Wang, Aisha Behr, Kai Zhao, Jeffrey Friedman, Xu Chu, and Amine Anoun. 2026. Adaptive memory admission control for llm agents. arXiv preprint arXiv:2603.04549

    arXiv 2026

  25. [25]

    Wanjun Zhong, Lianghong Guo, Qiqi Gao, He Ye, and Yanlin Wang. 2024. Memorybank: Enhancing large language models with long-term memory. In Proceedings of the AAAI conference on artificial intelligence, volume 38, pages 19724--19731

    2024

  26. [26]

    Wenhui Zhu, Xiwen Chen, Zhipeng Wang, Jingjing Wang, Xuanzhao Dong, Minzhou Huang, Rui Cai, Hejian Sang, Hao Wang, Peijie Qiu, and 1 others. 2026. Ariadnemem: Threading the maze of lifelong memory for llm agents. arXiv preprint arXiv:2603.03290

    arXiv 2026