Pith. sign in

REVIEW 4 major objections 5 minor 64 references

Hierarchical Graph Memory for LLM Agents with Path-level Localization and Rewrite

T0 review · 4 major / 5 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 A memory that organizes facts into hierarchical regions and rewrites only the affected evidence path outperforms flat memory agents on long-term QA and conflict benchmarks.

desk verdict HiGram's path-level localization is a genuinely new idea, but the core scoring functions are unspecified and the conflict-update mechanism has a plausible failure mode that the authors must resolve. read the letter →

arxiv 2608.05095 v1 pith:4NMKA5CO submitted 2026-08-05 cs.AI

Xiawei Yue , Boran Wang , Xiaoqing Zhang , Shuxin Zheng , Ziwei Zhang This is my paper
classification cs.AI
keywords hierarchicalgraphmemoryLLMagentsupdateevidencelocalizationpath-levelrewritelong-termreasoningconflict-awaretokenefficiency
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 tries to establish that an LLM agent's long-term memory should be organized hierarchically and updated along localized evidence paths, rather than stored as a flat graph of independently rewritten facts. It argues that answers depend on small connected chains of memory units, so retrieval and rewriting should first locate the relevant subgraph and the affected path, then update unit states and their dependencies together. The proposed HiGram system claims to achieve better answer quality and large token savings than baselines on long-term conversational question answering, and better handling of dynamic, static, and conditional memory conflicts. If correct, this points toward memory designs where evidence localization, not context expansion, is the main lever for both accuracy and cost.

What carries the argument

The central object is a hierarchical graph memory together with MicroGraph-based path-level localization and coordinated rewriting. A MicroGraph is a pair $(v^{sub}_{t,j}, v^{cat}_{t,j})$ of a subject node and an object-category node that defines a localized region of the global graph memory, containing the MemoryUnits, context nodes, and dependency edges associated with that subject-category pair. This object carries the argument because it lets the system first narrow retrieval to a support subgraph, then identify a single affected evidence path, and finally restrict all rewriting to that bounded path. The scoring functions $R(B, A_t)$ and $\phi_H(P, M_{temp}^t)$ select the MicroGraphs and the evidence path, and the update rule $G_{t+1} = \text{Update}(G_t, \text{Rewrite}_{inter}(\text{Rewrite}_{intra}(\hat{P}_t, M_{temp}^t), M_{temp}^t))$ performs the state and dependency updates in one coordinated pass.

What would settle it

Re-run the LoCoMo evaluation with the anchor extractor replaced by random selection of $K_g$ MicroGraphs among the candidates while keeping every other component fixed; if token-level F1 and LLM-judged scores stay near the reported HiGram numbers, then path-level localization is not doing the causal work, whereas a drop to baseline levels would confirm that it is.

Watch

Extended reading notes

Core claim

HiGram's central claim is that coupling coarse-to-fine memory organization with path-level localization and coordinated rewriting yields a memory that evolves consistently and cheaply. Concretely, it stores facts as MemoryUnits under upper-level subject, object-category, and context nodes; defines a MicroGraph as a localized region determined by a subject node and an object-category node; extracts anchors from the query and update to select the top-$K_g$ MicroGraphs; enumerates candidate memory paths within the resulting support subgraph; and scores each path for consistency with the temporary update units. The highest-scoring path defines the rewrite region, where intra-unit rewriting updates the internal states of affected MemoryUnits and inter-unit rewriting revises dependency edges so that outdated dependencies are excluded from the active evidence view. The paper reports that this method outperforms six baselines on LoCoMo in token-level F1, BLEU, and LLM-judged quality while using only 7.2% of the tokens consumed by full-context inference under GPT-5.4, and that it leads MemConflict on static conflict accuracy and on both evidence-selection metrics.

Load-bearing premise

The entire method depends on the anchor extractor and the two scoring functions correctly locating the support subgraph and the one evidence path that actually contains the affected facts; if they pick the wrong region, the coordinated rewrite misses the affected MemoryUnits even though the rewriting rules themselves are sound.

Editorial extensions

If this is right

  • Long-term question-answering agents can maintain or exceed full-context answer accuracy while spending a small fraction of the tokens, because retrieval, evidence selection, and rewriting are all confined to localized evidence paths.
  • Memory updates no longer require repeated unit-wise rewrites: a single coordinated pass over the affected evidence path updates both the internal facts and the dependency edges that connect them, which is the mechanism behind HiGram's token savings.
  • Conflict handling improves across dynamic, static, and conditional conflicts because superseded and pending units are retained as history while outdated dependencies are marked and excluded from the active evidence view.
  • Evidence selection quality improves alongside answer quality: on MemConflict, the gold evidence appears earlier in the retrieved ranking (higher SEH@3 and SRS), indicating that localization helps retrieval, not just generation.

Reading between the lines

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

  • A testable extension: the same path-level localization idea could transfer to knowledge editing in LLMs, where a single fact edit would be accompanied by a minimal evidence path rewrite to prevent cascading inconsistencies in derived conclusions.
  • A stress test implied by the design: on conversations with high entity fan-out (one subject linked to many object categories), the subject-category anchor pairing may become a bottleneck; a dataset engineered with such fan-out would reveal whether the fixed top-$K_g$ budget of 12 MicroGraphs can still locate the relevant regions.
  • If the claimed token savings come from localization rather than compression, then an exact evidence-coverage metric should show that the selected path contains more gold evidence per consumed token than flat retrieval; the paper's SEH@3 and SRS results are consistent with this but do not directly measure token-normalized coverage.
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

4 major / 5 minor

Summary. The paper proposes HiGram, an evolving hierarchical graph memory framework for LLM agents. Memory is organized as a coarse-to-fine graph with upper-level nodes and fine-grained MemoryUnits. Given a query and an update, HiGram constructs temporary MemoryUnits, extracts anchors, selects relevant MicroGraphs to form a support subgraph, and then chooses an evidence path via a scoring function phi_H. A coordinated rewriting step updates the internal states of affected MemoryUnits and their inter-unit dependencies within that localized path. The method is evaluated on LoCoMo and MemConflict, reporting improvements in answer quality, token efficiency, and conflict handling across GPT-5.4 and GPT-4o backbones.

Significance. If the reported results hold, HiGram addresses a real and important problem: maintaining long-term memory for LLM agents under continual updates while keeping evidence structures consistent and retrieval efficient. The paper's strengths include evaluation on external benchmarks, two LLM backbones, several strong baselines, ablation studies, and hyperparameter sensitivity analyses. The gains on MemConflict Static and Conditional conflicts are large, and the result that MicroGraph organization and support-subgraph localization both contribute is informative. However, the central mechanism is under-specified: the scoring functions that drive localization are never concretely defined, and the token-efficiency metric excludes memory-maintenance costs. The empirical claim of substantial improvement therefore cannot currently be fully audited, though the work is promising.

major comments (4)
  1. [Section 3.2, Eqs. (5)-(8)] The scoring functions R(B,A_t) and phi_H(P,M_temp^t) are described only qualitatively: R is said to measure 'subject matching and object-category compatibility' and phi_H to 'consider matching of MemoryUnit attributes, dependency consistency, temporal validity, and contextual compatibility.' Their concrete implementations are never given. Because Eq. (7) uses R to select the top-Kg MicroGraphs and Eq. (8) uses phi_H to select the affected evidence path, the entire localization pipeline—and the subsequent claim that rewriting operates on the affected path—is neither reproducible nor auditable. Please specify how these scores are computed, including any LLM prompts, exact attribute weights, or learned components, so a reader could reimplement the method.
  2. [Section 3.2, Eq. (8) and Section 3.3] For Static and Dynamic conflicts, the argmax formulation in Eq. (8) with phi_H described as 'consistency' may systematically down-rank the path containing the to-be-superseded fact. A temporary MemoryUnit asserting a new value for a known subject makes the old assertion inconsistent (different object, outdated temporal validity), so unrelated but non-conflicting paths could receive higher phi_H scores. If so, bP_t would exclude the affected MemoryUnit, and the coordinated rewriting in Section 3.3 would not revise the fact that must be superseded, contradicting the Table 2 conflict-handling gains. The paper must clarify how phi_H balances contradiction against subject/attribute matching and ideally provide an analysis or examples showing that conflict updates select evidence paths that actually contain the affected MemoryUnits.
  3. [Section 4.1, Token Length metric] The reported token efficiency is measured only for the final answer generation call, explicitly 'excluding offline memory construction and update.' HiGram's anchor extraction, path enumeration, and coordinated rewriting all consume additional LLM tokens or computation that are not counted. Therefore the claim that HiGram uses only 7.2% of full-context tokens (Table 1) applies only to the answer call, not to the total operational cost of the memory system. Please report the full maintenance and update budget, or qualify the token-efficiency claim to state that it covers answer generation only.
  4. [Tables 1 and 2] No variance or significance testing is reported. LLM-based evaluation metrics such as F1, BLEU, and LLM-as-Judge are noisy, and single-run differences of a few points may not be reliable. The headline improvements on LoCoMo (Table 1) and MemConflict (Table 2) should be accompanied by multiple runs with means and standard deviations, or at least significance tests, before the claim of 'substantial improvements' can be considered established.
minor comments (5)
  1. [Abstract and Introduction] The phrase 'our method demonstrate substantial improvements' should be 'our method demonstrates substantial improvements'; the same grammatical error appears in the Abstract.
  2. [Section 3.2, Eq. (8)] Equation (8) contains a stray closing parenthesis and a malformed expression: 'arg max_{P in P_cand^t} phi_H(P, M_temp^t))' has an extra parenthesis after M_temp^t.
  3. [Table 1] There are duplicated numbers in the table: the LoCoMo row under GPT-5.4 shows '74.89 74.89' for average F1/BLEU/LLM-J columns, and the A-MEM row under GPT-4o shows '97.31 97.31'; these should be corrected.
  4. [Section 3.2] The paper refers to 'evidence path' as a connected chain of MemoryUnits, but does not formally define whether context nodes or upper-level nodes can appear in a path, nor how paths are enumerated or deduplicated when Kp candidate paths are formed.
  5. [Section 4.1] The implementation details state a default path depth h=3, but h does not appear in any equation in Section 3; please clarify how path depth is used during candidate path enumeration.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: HiGram is an empirical memory system evaluated on external benchmarks, with no fitted parameter renamed as a prediction and no load-bearing self-citation chain.

full rationale

HiGram is an empirical memory-management system, not a derivation of predictions from fitted constants. Its central claims, improved answer quality, token efficiency, and conflict handling, are supported by comparisons against external baselines on LoCoMo and MemConflict (Tables 1 and 2), plus ablations and hyperparameter sensitivity analyses. The path-level localization in Eq. (8) selects the affected evidence path by maximizing the scoring function phi_H, which is described as measuring consistency between candidate paths and temporary MemoryUnits; even if this scoring function has defects, such as down-ranking the path containing a superseded fact under conflict corrections, that is a correctness risk rather than circularity, because the empirical results are not forced by the definition of phi_H. No fitted parameter is renamed as a prediction: the retrieved-region size K_g and evidence-path budget K_p are hyperparameters whose effects are reported in Figure 3, and the method is not trained or calibrated on the benchmark gold evidence. The evaluation uses an LLM judge from the same model family as the answer generator, which is an evaluation bias, not a circular derivation. There is also no load-bearing self-citation chain: the related work and benchmark citations are external, and the paper does not invoke a uniqueness theorem or prior result by the same authors to justify its design choices. Consequently, the claimed improvements are externally falsifiable claims about a concrete system and do not reduce by construction to the paper's inputs.

Assumptions & free parameters 3 free parameters · 4 assumptions · 4 invented entities

The central claim rests on LLM-based extraction and scoring assumptions that are not specified or separately validated, plus hand-set retrieval budgets (Kg, Kp, path depth). The system components (MemoryUnit, MicroGraph, statuses) are internal constructs rather than independently evidenced entities.

free parameters (3)
  • Kg = 12
    Number of MicroGraphs retrieved; hand-set in Section 4.1 and varied in sensitivity analysis (Figure 3).
  • Kp = 24
    Evidence path enumeration budget; hand-set in Section 4.1 and varied in Figure 3.
  • path depth h = 3
    Default evidence path depth; chosen in implementation details without a selection procedure.
assumptions (4)
  • domain assumption LLM can reliably extract anchors and build temporary MemoryUnits from queries and updates.
    Invoked in Section 3.2 through anch(M_temp); the method depends on accurate extraction but the paper provides no validation of extraction quality.
  • domain assumption LLM-based relevance score R and path score phi_H correctly rank evidence.
    Equations 6-8 define these scores only notationally; their implementation is not specified, so correctness is assumed.
  • domain assumption MemoryUnit schema and the active/superseded/outdated/pending statuses are sufficient to represent conflict and temporal validity.
    The coordinated rewriting in Section 3.3 relies on these statuses to preserve or invalidate evidence; no independent check that the schema captures all relevant conflict types.
  • domain assumption LLM-as-judge scores are a valid measure of answer quality.
    LLM-J is reported as a semantic correctness metric; potential judge bias is not analyzed.
invented entities (4)
  • MemoryUnit
    purpose: Atomic memory record storing subject, relation, object, object category, transaction time, context, confidence, and status.
    Internal data structure of HiGram; no evidence outside the paper's system-level benchmarks.
  • MicroGraph
    purpose: Localized subgraph defined by a subject/object-category pair used to constrain retrieval before path selection.
    A conceptual grouping introduced by the paper; its usefulness is only indirectly supported by ablation.
  • MemoryUnit status (active/superseded/outdated/pending)
    purpose: State machine that marks whether a fact participates in current retrieval or is historical.
    Internal bookkeeping device; the validity of the status transitions is not independently tested.
  • Temporary MemoryUnit
    purpose: Query/update-derived memory records that are excluded from the graph until committed by rewriting.
    Part of the proposed localization pipeline; no external handle.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Hierarchical Graph Memory for LLM Agents with Path-level Localization and Rewrite." pith.science (2026). https://pith.science/paper/4NMKA5CO

@misc{pith2026260805095,
  author       = {Pith},
  title        = {Pith review of: Hierarchical Graph Memory for LLM Agents with Path-level Localization and Rewrite},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4NMKA5CO}},
  note         = {Machine review of arXiv:2608.05095}
}
read the original abstract

Agents for long term reasoning require a memory that can be efficiently and effectively updated over time, as new facts and external feedback continue to arrive. Recently, graph memory has been adopted to offer structural organization for multi-hop retrieval and reasoning. However, existing methods store all memories in a flat graph, and accumulated historical memories can introduce irrelevant contexts and increase the cost of evidence selection during retrieval. Moreover, they typically update memory units independently, requiring repeated unit-wise rewrite to cover related changes. To address these issues, we propose HiGram, an evolving hierarchical graph memory framework with path-level localization and rewriting. Specifically, we first propose a hierarchical graph memory, which organizes the memory into coarse-to-fine architecture composed of upper-level nodes and MemoryUnits, thereby reducing the amount of irrelevant information during retrieval. We further propose MicroGraph-based path-level localization, which leverages query and update conditioned MicroGraphs to identify support subgraph and evidence path before rewrite. Finally, we propose a coordinated rewriting method that jointly revises intra-unit memory and inter-unit dependencies, enable valid dependency structures updating in the localized evidence path. Experiments on benchmarks for long-term conversational question answering and conflict-aware memory evaluation demonstrate that our method demonstrate substantial improvements over baselines in answer quality and token efficiency. Besides, our method improves answer accuracy and query-valid evidence selection under dynamic, static, and conditional conflicts.

Figures

Figures reproduced from arXiv: 2608.05095 by the authors.

Figure 1
Figure 1. An overview of HiGram. Our method organizes memory into a hierarchical graph structure with upper-level nodes connecting MemoryUnits that store factual information. Then, given a query and an update, our MicroGraph-based path-level localization module retrieves relevant memory MicroGraphs to build a support subgraph, and identifies the affected evidence path. Lastly, coordinated rewriting updates MemoryUnit states a… view at source ↗
Figure 2
Figure 2. Results of ablation of Memory Organization and Evidence Localization on LoCoMo. [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Sensitivity analysis of key hyperparameters on Lo [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

64 extracted references · 27 canonical work pages

  1. [1]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics , pages=

    Evaluating very long-term conversational memory of llm agents , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics , pages=

  2. [2]

    Proceedings of the AAAI conference on artificial intelligence , number=

    Memorybank: Enhancing large language models with long-term memory , author=. Proceedings of the AAAI conference on artificial intelligence , number=

    work page

  3. [3]

    Advances in Neural Information Processing Systems , volume=

    A-mem: Agentic memory for llm agents , author=. Advances in Neural Information Processing Systems , volume=

  4. [4]

    2023 , journal=

    MemGPT: towards LLMs as operating systems , author=. 2023 , journal=

    work page 2023

  5. [5]

    International Conference on Machine Learning , pages=

    A Human-Inspired Reading Agent with Gist Memory of Very Long Contexts , author=. International Conference on Machine Learning , pages=. 2024 , organization=

    2024

  6. [6]

    arXiv preprint arXiv:2504.19413 , year=

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

    arXiv

  7. [7]

    arXiv preprint arXiv:2501.13956 , year=

    Zep: a temporal knowledge graph architecture for agent memory , author=. arXiv preprint arXiv:2501.13956 , year=

    arXiv

  8. [8]

    MemConflict: Evaluating Long-Term Memory Systems Under Memory Conflicts

    MemConflict: Evaluating Long-Term Memory Systems Under Memory Conflicts , author=. arXiv preprint arXiv:2605.20926 , year=

    work page Pith review arXiv

Show all 64 references
  1. [9]

    The Thirteenth International Conference on Learning Representations , year=

    LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory , author=. The Thirteenth International Conference on Learning Representations , year=

  2. [10]

    Advances in Neural Information Processing Systems , volume=

    Augmenting language models with long-term memory , author=. Advances in Neural Information Processing Systems , volume=

  3. [11]

    Proceedings of the 41st International Conference on Machine Learning , pages=

    MEMORYLLM: towards self-updatable large language models , author=. Proceedings of the 41st International Conference on Machine Learning , pages=

  4. [12]

    arXiv preprint arXiv:2308.08239 , year=

    Memochat: Tuning llms to use memos for consistent long-range open-domain conversation , author=. arXiv preprint arXiv:2308.08239 , year=

    arXiv

  5. [13]

    Findings of the Association for Computational Linguistics , pages=

    Long time no see! open-domain conversation with long-term persona memory , author=. Findings of the Association for Computational Linguistics , pages=

  6. [14]

    Proceedings of the 36th annual acm symposium on user interface software and technology , pages=

    Generative agents: Interactive simulacra of human behavior , author=. Proceedings of the 36th annual acm symposium on user interface software and technology , pages=

  7. [15]

    Advances in neural information processing systems , volume=

    Reflexion: Language agents with verbal reinforcement learning , author=. Advances in neural information processing systems , volume=

  8. [16]

    arXiv preprint arXiv:2311.08719 , year=

    Think-in-memory: Recalling and post-thinking enable llms with long-term memory , author=. arXiv preprint arXiv:2311.08719 , year=

    arXiv

  9. [17]

    Proceedings of the 31st International Conference on Computational Linguistics , pages=

    Personalized large language model assistant with evolving conditional memory , author=. Proceedings of the 31st International Conference on Computational Linguistics , pages=

  10. [18]

    ICML 2025 Workshop on Long-Context Foundation Models , year=

    Evaluating memory in llm agents via incremental multi-turn interactions , author=. ICML 2025 Workshop on Long-Context Foundation Models , year=

    2025

  11. [19]

    Findings of the Association for Computational Linguistics: ACL 2025 , pages=

    Membench: Towards more comprehensive evaluation on the memory of llm-based agents , author=. Findings of the Association for Computational Linguistics: ACL 2025 , pages=

    2025

  12. [20]

    arXiv preprint arXiv:2511.03506 , year=

    Halumem: Evaluating hallucinations in memory systems of agents , author=. arXiv preprint arXiv:2511.03506 , year=

  13. [21]

    arXiv preprint arXiv:2602.01313 , year=

    Evaluating long-horizon memory for multi-party collaborative dialogues , author=. arXiv preprint arXiv:2602.01313 , year=

  14. [22]

    Advances in neural information processing systems , volume=

    Retrieval-augmented generation for knowledge-intensive nlp tasks , author=. Advances in neural information processing systems , volume=

  15. [23]

    Proceedings of the 2020 conference on empirical methods in natural language processing , pages=

    Dense passage retrieval for open-domain question answering , author=. Proceedings of the 2020 conference on empirical methods in natural language processing , pages=

    2020

  16. [24]

    International conference on machine learning , pages=

    Retrieval augmented language model pre-training , author=. International conference on machine learning , pages=

  17. [25]

    International conference on machine learning , pages=

    Improving language models by retrieving from trillions of tokens , author=. International conference on machine learning , pages=

  18. [26]

    International conference on learning representations , volume=

    Self-rag: Learning to retrieve, generate, and critique through self-reflection , author=. International conference on learning representations , volume=

  19. [27]

    Proceedings of the 2023 conference on empirical methods in natural language processing , pages=

    Active retrieval augmented generation , author=. Proceedings of the 2023 conference on empirical methods in natural language processing , pages=

    2023

  20. [28]

    Proceedings of the 61st annual meeting of the association for computational linguistics , pages=

    Interleaving retrieval with chain-of-thought reasoning for knowledge-intensive multi-step questions , author=. Proceedings of the 61st annual meeting of the association for computational linguistics , pages=

  21. [29]

    Transactions of the association for computational linguistics , volume=

    Lost in the middle: How language models use long contexts , author=. Transactions of the association for computational linguistics , volume=

  22. [30]

    arXiv preprint arXiv:2404.16130 , year=

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

    arXiv

  23. [31]

    Findings of the Association for Computational Linguistics: EMNLP 2025 , year=

    Lightrag: Simple and fast retrieval-augmented generation , author=. Findings of the Association for Computational Linguistics: EMNLP 2025 , year=

    2025

  24. [32]

    Findings of the Association for Computational Linguistics: EMNLP 2024 , pages=

    Graphreader: Building graph-based agent to enhance long-context abilities of large language models , author=. Findings of the Association for Computational Linguistics: EMNLP 2024 , pages=

    2024

  25. [33]

    International Conference on Learning Representations , volume=

    Raptor: Recursive abstractive processing for tree-organized retrieval , author=. International Conference on Learning Representations , volume=

  26. [34]

    Advances in neural information processing systems , volume=

    Hipporag: Neurobiologically inspired long-term memory for large language models , author=. Advances in neural information processing systems , volume=

  27. [35]

    Proceedings of the Thirty-Fourth International Joint Conference on Artificial Intelligence , pages=

    AriGraph: learning knowledge graph world models with episodic memory for LLM agents , author=. Proceedings of the Thirty-Fourth International Joint Conference on Artificial Intelligence , pages=

  28. [36]

    Proceedings of the AAAI conference on artificial intelligence , number=

    Graph of thoughts: Solving elaborate problems with large language models , author=. Proceedings of the AAAI conference on artificial intelligence , number=

  29. [37]

    Advances in neural information processing systems , volume=

    Tree of thoughts: Deliberate problem solving with large language models , author=. Advances in neural information processing systems , volume=

  30. [38]

    Advances in neural information processing systems , volume=

    Chain-of-thought prompting elicits reasoning in large language models , author=. Advances in neural information processing systems , volume=

  31. [39]

    arXiv preprint arXiv:2201.08236 , year=

    Temporal knowledge graph completion: A survey , author=. arXiv preprint arXiv:2201.08236 , year=

    arXiv

  32. [40]

    international conference on machine learning , pages=

    Know-evolve: Deep temporal reasoning for dynamic knowledge graphs , author=. international conference on machine learning , pages=

  33. [41]

    Communications of the ACM , volume=

    Maintaining knowledge about temporal intervals , author=. Communications of the ACM , volume=. 1983 , publisher=

    1983

  34. [42]

    TIMEDIAL: Temporal commonsense reasoning in dialog , author=. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers) , pages=

  35. [43]

    Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , pages=

    Knowledge conflicts for llms: A survey , author=. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , pages=

    2024

  36. [44]

    arXiv preprint arXiv:2310.00935 , year=

    Resolving knowledge conflicts in large language models , author=. arXiv preprint arXiv:2310.00935 , year=

    arXiv

  37. [45]

    Findings of the Association for Computational Linguistics: EMNLP 2024 , pages=

    Who’s who: Large language models meet knowledge conflicts in practice , author=. Findings of the Association for Computational Linguistics: EMNLP 2024 , pages=

    2024

  38. [46]

    Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pages=

    Factscore: Fine-grained atomic evaluation of factual precision in long form text generation , author=. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pages=

    2023

  39. [47]

    International Conference on Learning Representations , year=

    Bertscore: Evaluating text generation with bert , author=. International Conference on Learning Representations , year=

  40. [48]

    Proceedings of the 40th annual meeting of the Association for Computational Linguistics , pages=

    Bleu: a method for automatic evaluation of machine translation , author=. Proceedings of the 40th annual meeting of the Association for Computational Linguistics , pages=

  41. [49]

    Text summarization branches out , pages=

    Rouge: A package for automatic evaluation of summaries , author=. Text summarization branches out , pages=

  42. [50]

    Communications of the ACM , volume=

    Open information extraction from the web , author=. Communications of the ACM , volume=. 2008 , publisher=

    2008

  43. [51]

    2025 , month = feb, howpublished =

    2025

  44. [52]

    Letta: A Platform for Stateful Agents with Persistent Memory , year =

  45. [53]

    arXiv preprint arXiv:2507.03724 , year=

    Memos: A memory os for ai system , author=. arXiv preprint arXiv:2507.03724 , year=

    arXiv

  46. [54]

    International Conference on Learning Representations , year=

    ReAct: Synergizing Reasoning and Acting in Language Models , author=. International Conference on Learning Representations , year=

  47. [55]

    International Conference on Learning Representations , year=

    Least-to-Most Prompting Enables Complex Reasoning in Large Language Models , author=. International Conference on Learning Representations , year=

  48. [56]

    arXiv preprint arXiv:2507.22925 , year=

    Hierarchical Memory for High-Efficiency Long-Term Reasoning in LLM Agents , author=. arXiv preprint arXiv:2507.22925 , year=

    arXiv

  49. [57]

    Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics , pages=

    Gam: Hierarchical graph-based agentic memory for llm agents , author=. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics , pages=

  50. [58]

    Advances in Neural Information Processing Systems , volume=

    G-memory: Tracing hierarchical memory for multi-agent systems , author=. Advances in Neural Information Processing Systems , volume=

  51. [59]

    Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    Memory os of ai agent , author=. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    2025

  52. [60]

    The eleventh international conference on learning representations , year=

    Mass-editing memory in a transformer , author=. The eleventh international conference on learning representations , year=

  53. [61]

    Advances in neural information processing systems , year=

    Locating and editing factual associations in gpt , author=. Advances in neural information processing systems , year=

  54. [62]

    Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , pages=

    Lifelong knowledge editing for llms with retrieval-augmented continuous prompt learning , author=. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , pages=

    2024

  55. [63]

    International Conference on Machine Learning , pages=

    Memory-based model editing at scale , author=. International Conference on Machine Learning , pages=

  56. [64]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    History matters: Temporal knowledge editing in large language model , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

Pith tools

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