pith. sign in

arxiv: 2606.28349 · v1 · pith:4W5TEPHAnew · submitted 2026-06-03 · 💻 cs.IR · cs.AI

HMARS: A Hierarchical Multi-Agent Memory System for Long-Context Reasoning

Zeju Li , Ziyang Zheng , Yizhou Zhou , Qiang Xu This is my paper

Pith reviewed 2026-06-30 11:28 UTC · model grok-4.3

classification 💻 cs.IR cs.AI
keywords long-context reasoningmulti-agent systemsmemory managementretrieval-augmented generationevidence retrievalhierarchical agentscontext-dependent relevance
0
0 comments X

The pith

A hierarchical multi-agent system manages long contexts by assigning sub-agents to bounded memory regions and mid-agents to query-specific coordination, retrieving supporting evidence more completely than top-K retrieval.

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

Standard retrieval-augmented generation reduces long-context reasoning to top-K chunk retrieval, which can discard relevant evidence when relevance depends on broader context. The paper proposes treating long contexts as managed memory instead, with sub-agents maintaining grounded access to bounded regions, mid-agents providing regional context and query coordination, and a frontier model performing final reasoning over retrieved evidence pages. It introduces two diagnostic benchmarks for evidence breadth and context-dependent relevance. On long-document and multi-turn memory tasks, this structure outperforms retrieval, reranking, full-context, graph-based, and agentic baselines, with gains traced to more complete evidence coverage rather than prompt changes.

Core claim

HMARS achieves the best overall performance across long-document and multi-turn memory tasks against retrieval, reranking, full-context, graph-based, and agentic long-context baselines, with evidence coverage analysis showing the gains come from retrieving the required supporting evidence more completely.

What carries the argument

The hierarchical multi-agent memory system, in which sub-agents maintain access to bounded memory regions, mid-agents manage regional context and query-specific coordination, and a frontier model performs final reasoning over evidence pages.

If this is right

  • The system outperforms standard retrieval, reranking, full-context, graph-based, and agentic baselines on long-document and multi-turn tasks.
  • Performance gains trace directly to more complete retrieval of supporting evidence rather than changes to the final reasoning prompt.
  • The approach targets evidence breadth and context-dependent relevance through its diagnostic benchmarks.
  • Sub-agents and mid-agents together reduce passive access losses that occur when relevance depends on surrounding context.

Where Pith is reading between the lines

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

  • The structure might allow memory regions to be updated or reorganized dynamically without re-running full retrieval.
  • Similar coordination layers could be tested on tasks that combine multiple long documents where evidence spans document boundaries.
  • If the hierarchy generalizes, it could reduce the need for ever-larger context windows by focusing agent effort on relevant sub-regions.

Load-bearing premise

Dividing long contexts into hierarchical agent-managed regions will surface context-dependent relevant evidence without the information loss that occurs in flat top-K retrieval.

What would settle it

A diagnostic task in which flat top-K retrieval covers all required evidence but the hierarchical boundaries cause HMARS to miss some pieces, producing lower performance than the flat baseline.

Figures

Figures reproduced from arXiv: 2606.28349 by Qiang Xu, Yizhou Zhou, Zeju Li, Ziyang Zheng.

Figure 1
Figure 1. Figure 1: Passive RAG chunks documents, embeds chunks independently, and retrieves top-K evidence, but may miss context-dependent evidence. HMARS treats long inputs as managed memory, where sub-agents as￾sess local regions, mid-agents coordinate shared context, and a frontier model performs final reasoning. top-K, and passes them to a generator. This passive interface is efficient, but assumes relevance is local and… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of HMARS. Indexing phase: The document is partitioned into local page-level regions; sub-agents build local memory cards, which mid-agents aggregate into shared regional memory. Query phase: Mid-agents broadcast query interpretations and sub-agent-specific hints; all sub-agents select relevant pages in parallel, and the base agent synthesizes the final answer from the ordered grounded evidence. Pr… view at source ↗
Figure 3
Figure 3. Figure 3: Architectural ablations for HMARS. Mid-agents provide shared regional context. Removing the mid-agent layer drops overall ac￾curacy from 0.838 to 0.748, with the largest degra￾dation on multi-turn memory. This indicates that sub-agents should not operate as isolated local read￾ers. The mid-agent aggregates sub-agent memories into a shared regional view and broadcasts query￾specific guidance, helping local … view at source ↗
Figure 5
Figure 5. Figure 5: Accuracy–compute trade-off on the full evalu [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 4
Figure 4. Figure 4: Granularity ablations for HMARS. Page size P controls the evidence unit returned to the base model, while sub-agent capacity Csub controls the size of each local memory region. 4.7 Ablation on Model Allocation Finally, we test whether local memory management requires a larger sub-agent model, results are shown in [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
read the original abstract

Long-context reasoning requires models to access, retrieve, and integrate evidence scattered across documents, dialogues, and accumulated interaction histories. Standard retrieval-augmented generation reduces this problem to top-$K$ chunk retrieval, but such passive access can discard relevant evidence before reasoning begins, especially when relevance depends on broader context. We propose HMARS, a hierarchical multi-agent memory system that treats long contexts as managed memory rather than a flat retrieval corpus. Sub-agents maintain grounded access to bounded memory regions, mid-agents manage regional context and provide query-specific coordination, and a frontier model performs final reasoning over retrieved evidence pages. To evaluate this view, we construct two diagnostic benchmarks targeting evidence breadth and context-dependent relevance. Across long-document and multi-turn memory tasks, HMARS achieves the best overall performance against retrieval, reranking, full-context, graph-based, and agentic long-context baselines. Evidence coverage analysis further shows that its gains come from retrieving the required supporting evidence more completely, rather than merely changing the final answer prompt.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 0 minor

Summary. The paper introduces HMARS, a hierarchical multi-agent memory system for long-context reasoning in which sub-agents maintain grounded access to bounded memory regions, mid-agents manage regional context and query-specific coordination, and a frontier model performs final reasoning over retrieved evidence pages. It constructs two diagnostic benchmarks targeting evidence breadth and context-dependent relevance, and claims that HMARS achieves the best overall performance against retrieval, reranking, full-context, graph-based, and agentic baselines on long-document and multi-turn memory tasks, with gains attributable to more complete retrieval of supporting evidence.

Significance. If the performance claims and evidence-coverage analysis are substantiated with reproducible experiments, the work would offer a concrete advance over passive top-K retrieval by showing how hierarchical agent coordination can mitigate context-dependent information loss in long-context settings.

major comments (2)
  1. [Abstract] Abstract: the assertion that 'HMARS achieves the best overall performance' and that 'its gains come from retrieving the required supporting evidence more completely' is presented without any quantitative metrics, result tables, statistical tests, baseline implementation details, or benchmark construction information, rendering the central empirical claim impossible to evaluate.
  2. [Abstract] Abstract: the motivating claim that the hierarchical division into sub-agents and mid-agents 'will reliably surface context-dependent relevant evidence without the information loss that occurs in flat top-K retrieval' is unsupported by any description of memory-bounding mechanics, coordination protocols, evidence-page construction, or the precise evidence-coverage metric.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on the abstract. We agree that the abstract can be strengthened to better support its empirical claims and mechanistic descriptions while remaining concise, and we will revise it accordingly. The main body of the manuscript already contains the requested details on metrics, baselines, benchmarks, and system mechanics.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the assertion that 'HMARS achieves the best overall performance' and that 'its gains come from retrieving the required supporting evidence more completely' is presented without any quantitative metrics, result tables, statistical tests, baseline implementation details, or benchmark construction information, rendering the central empirical claim impossible to evaluate.

    Authors: We acknowledge that the abstract, as a high-level summary, does not embed the full quantitative details. The manuscript provides these elements in the experimental evaluation, including result tables comparing against all listed baselines, statistical tests, baseline implementation details, and benchmark construction information. To address the concern directly, we will revise the abstract to incorporate key quantitative highlights (such as overall performance gains and evidence coverage improvements) and explicit references to the supporting tables and sections. revision: yes

  2. Referee: [Abstract] Abstract: the motivating claim that the hierarchical division into sub-agents and mid-agents 'will reliably surface context-dependent relevant evidence without the information loss that occurs in flat top-K retrieval' is unsupported by any description of memory-bounding mechanics, coordination protocols, evidence-page construction, or the precise evidence-coverage metric.

    Authors: The abstract summarizes the approach at a high level. The full manuscript details the memory-bounding mechanics for sub-agents, the coordination protocols used by mid-agents, evidence-page construction, and the precise evidence-coverage metric in the methods and evaluation sections. We will revise the abstract to include a brief description of these components to better ground the motivating claim. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical performance claims with no derivations

full rationale

The paper describes an architectural proposal (hierarchical sub-agents, mid-agents, evidence pages) and reports empirical results on constructed benchmarks against baselines. No equations, derivations, fitted parameters, or mathematical claims appear in the abstract or described full text. Central claims rest on observed performance and evidence-coverage metrics rather than any reduction to self-definitions, self-citations, or renamed inputs. The evaluation is self-contained as an experimental comparison without load-bearing self-referential steps.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Only the abstract is available; no free parameters, axioms, or invented entities are identifiable from the provided text.

pith-pipeline@v0.9.1-grok · 5708 in / 1139 out tokens · 44023 ms · 2026-06-30T11:28:28.087933+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

85 extracted references · 38 canonical work pages · 15 internal anchors

  1. [2]

    International Conference on Learning Representations , volume=

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

  2. [5]

    arXiv preprint arXiv:2401.11504 , year=

    With greater text comes greater necessity: Inference-time training helps long text generation , author=. arXiv preprint arXiv:2401.11504 , year=

    work page arXiv

  3. [6]

    LLMLingua: Compressing prompts for accelerated inference of large language models.arXiv preprint arXiv:2310.05736, 2023

    Llmlingua: Compressing prompts for accelerated inference of large language models , author=. arXiv preprint arXiv:2310.05736 , year=

    work page arXiv

  4. [7]

    arXiv preprint arXiv:2403.12968 , year=

    Llmlingua-2: Data distillation for efficient and faithful task-agnostic prompt compression , author=. arXiv preprint arXiv:2403.12968 , year=

    work page arXiv

  5. [8]

    2025 , eprint=

    Context Cascade Compression: Exploring the Upper Limits of Text Compression , author=. 2025 , eprint=

    2025

  6. [9]

    Advances in Neural Information Processing Systems , volume=

    Learning to compress prompts with gist tokens , author=. Advances in Neural Information Processing Systems , volume=

  7. [10]

    arXiv preprint arXiv:2310.04408 , year=

    Recomp: Improving retrieval-augmented lms with compression and selective augmentation , author=. arXiv preprint arXiv:2310.04408 , year=

    work page arXiv

  8. [11]

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

    Compressing context to enhance inference efficiency of large language models , author=. Proceedings of the 2023 conference on empirical methods in natural language processing , pages=

    2023

  9. [12]

    arXiv preprint arXiv:2401.03462 , year=

    Long context compression with activation beacon , author=. arXiv preprint arXiv:2401.03462 , year=

    work page arXiv

  10. [13]

    arXiv preprint arXiv:2209.15189 , year=

    Learning by distilling context , author=. arXiv preprint arXiv:2209.15189 , year=

    work page arXiv

  11. [14]

    Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    A silver bullet or a compromise for full attention? a comprehensive study of gist token-based context compression , author=. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  12. [15]

    arXiv preprint arXiv:2405.03085 , year=

    Compressing long context for enhancing rag with amr-based concept distillation , author=. arXiv preprint arXiv:2405.03085 , year=

    work page arXiv

  13. [16]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Longllmlingua: Accelerating and enhancing llms in long context scenarios via prompt compression , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  14. [17]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Leveraging attention to effectively compress prompts for long-context llms , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  15. [18]

    arXiv preprint arXiv:2407.01527 , year=

    Kv cache compression, but what must we give in return? a comprehensive benchmark of long context capable approaches , author=. arXiv preprint arXiv:2407.01527 , year=

    work page arXiv

  16. [19]

    Perception Compressor: A Training-Free Prompt Compression Framework in Long Context Scenarios

    Tang, Jiwei and Xu, Jin and Lu, Tingwei and Zhang, Zhicheng and YimingZhao, YimingZhao and LinHai, LinHai and Zheng, Hai-Tao. Perception Compressor: A Training-Free Prompt Compression Framework in Long Context Scenarios. Findings of the Association for Computational Linguistics: NAACL 2025. 2025. doi:10.18653/v1/2025.findings-naacl.229

    work page doi:10.18653/v1/2025.findings-naacl.229 2025

  17. [20]

    Test-time learning for large language models,

    Test-Time Learning for Large Language Models , author=. arXiv preprint arXiv:2505.20633 , year=

    work page arXiv

  18. [21]

    Test-time training on nearest neighbors for large lan- guage models.arXiv preprint arXiv:2305.18466, 2023

    Test-time training on nearest neighbors for large language models , author=. arXiv preprint arXiv:2305.18466 , year=

    work page arXiv

  19. [22]

    arXiv preprint arXiv:2410.08020 , year=

    Efficiently learning at test-time: Active fine-tuning of llms , author=. arXiv preprint arXiv:2410.08020 , year=

    work page arXiv

  20. [23]

    2024 , eprint=

    StreamAdapter: Efficient Test Time Adaptation from Contextual Streams , author=. 2024 , eprint=

    2024

  21. [24]

    2024 , eprint=

    Generative Adapter: Contextualizing Language Models in Parameters with A Single Forward Pass , author=. 2024 , eprint=

    2024

  22. [25]

    2025 , eprint=

    InfiniteICL: Breaking the Limit of Context Window Size via Long Short-term Memory Transformation , author=. 2025 , eprint=

    2025

  23. [26]

    2021 , eprint=

    A General Language Assistant as a Laboratory for Alignment , author=. 2021 , eprint=

    2021

  24. [27]

    2022 , eprint=

    Prompt Injection: Parameterization of Fixed Inputs , author=. 2022 , eprint=

    2022

  25. [28]

    2022 , eprint=

    Prompt Compression and Contrastive Conditioning for Controllability and Toxicity Reduction in Language Models , author=. 2022 , eprint=

    2022

  26. [29]

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

    Adapting Language Models to Compress Contexts , author=. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pages=

    2023

  27. [30]

    In-context autoencoder for context compression in a large language model,

    In-context autoencoder for context compression in a large language model , author=. arXiv preprint arXiv:2307.06945 , year=

    work page arXiv

  28. [31]

    2024 , eprint=

    Compressed Context Memory For Online Language Model Interaction , author=. 2024 , eprint=

    2024

  29. [32]

    Advances in Neural Information Processing Systems , volume=

    H2o: Heavy-hitter oracle for efficient generative inference of large language models , author=. Advances in Neural Information Processing Systems , volume=

  30. [33]

    Advances in Neural Information Processing Systems , volume=

    Snapkv: Llm knows what you are looking for before generation , author=. Advances in Neural Information Processing Systems , volume=

  31. [34]

    Proceedings of the 39th International Conference on Machine Learning , pages =

    Efficient Test-Time Model Adaptation without Forgetting , author =. Proceedings of the 39th International Conference on Machine Learning , pages =. 2022 , editor =

    2022

  32. [35]

    2024 , eprint=

    Test-Time Model Adaptation with Only Forward Passes , author=. 2024 , eprint=

    2024

  33. [36]

    Tent: Fully Test-time Adaptation by Entropy Minimization

    Tent: Fully test-time adaptation by entropy minimization , author=. arXiv preprint arXiv:2006.10726 , year=

    work page internal anchor Pith review Pith/arXiv arXiv 2006

  34. [37]

    Know What You Don't Know: Unanswerable Questions for SQuAD

    Know what you don't know: Unanswerable questions for SQuAD , author=. arXiv preprint arXiv:1806.03822 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  35. [38]

    2021 , eprint=

    BookSum: A Collection of Datasets for Long-form Narrative Summarization , author=. 2021 , eprint=

    2021

  36. [39]

    ArXiv , year=

    Don't Give Me the Details, Just the Summary! Topic-Aware Convolutional Neural Networks for Extreme Summarization , author=. ArXiv , year=

  37. [40]

    2023 , eprint=

    GPQA: A Graduate-Level Google-Proof Q&A Benchmark , author=. 2023 , eprint=

    2023

  38. [41]

    Hashimoto , title =

    Rohan Taori and Ishaan Gulrajani and Tianyi Zhang and Yann Dubois and Xuechen Li and Carlos Guestrin and Percy Liang and Tatsunori B. Hashimoto , title =. GitHub repository , howpublished =. 2023 , publisher =

    2023

  39. [42]

    Transactions of the Association for Computational Linguistics , volume=

    Coqa: A conversational question answering challenge , author=. Transactions of the Association for Computational Linguistics , volume=. 2019 , publisher=

    2019

  40. [43]

    The Llama 3 Herd of Models

    The llama 3 herd of models , author=. arXiv preprint arXiv:2407.21783 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  41. [44]

    arXiv preprint arXiv:2510.00636 , year=

    Expected Attention: KV Cache Compression by Estimating Attention from Future Queries Distribution , author=. arXiv preprint arXiv:2510.00636 , year=

    work page arXiv

  42. [45]

    arXiv preprint arXiv:2502.00299 , year=

    Chunkkv: Semantic-preserving kv cache compression for efficient long-context llm inference , author=. arXiv preprint arXiv:2502.00299 , year=

    work page arXiv

  43. [46]

    500x C ompressor: Generalized Prompt Compression for Large Language Models

    Li, Zongqian and Su, Yixuan and Collier, Nigel. 500x C ompressor: Generalized Prompt Compression for Large Language Models. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025. doi:10.18653/v1/2025.acl-long.1219

    work page doi:10.18653/v1/2025.acl-long.1219 2025

  44. [48]

    arXiv e-prints , year = 2017, eid =

    triviaqa: A Large Scale Distantly Supervised Challenge Dataset for Reading Comprehension. arXiv e-prints , year = 2017, eid =

    2017

  45. [49]

    Training Verifiers to Solve Math Word Problems

    Training Verifiers to Solve Math Word Problems , author=. arXiv preprint arXiv:2110.14168 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  46. [52]

    International conference on learning representations , volume=

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

  47. [53]

    Advances in Neural Information Processing Systems , volume=

    Chain of agents: Large language models collaborating on long-context tasks , author=. Advances in Neural Information Processing Systems , volume=

  48. [55]

    Advances in Neural Information Processing Systems , volume=

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

  49. [58]

    Advances in neural information processing systems , volume=

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

  50. [59]

    Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Longbench v2: Towards deeper understanding and reasoning on realistic long-context multitasks , author=. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  51. [60]

    Proceedings of the 62nd annual meeting of the association for computational linguistics (volume 1: Long papers) , pages=

    Longbench: A bilingual, multitask benchmark for long context understanding , author=. Proceedings of the 62nd annual meeting of the association for computational linguistics (volume 1: Long papers) , pages=

  52. [61]

    Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    H-mem: Hierarchical memory for high-efficiency long-term reasoning in llm agents , author=. Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  53. [62]

    2026 , eprint=

    Agentic Retrieval-Augmented Generation: A Survey on Agentic RAG , author=. 2026 , eprint=

    2026

  54. [63]

    2025 , eprint=

    Hierarchical Document Refinement for Long-context Retrieval-augmented Generation , author=. 2025 , eprint=

    2025

  55. [67]

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

    LONGAGENT: achieving question answering for 128k-token-long documents through multi-agent collaboration , author=. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , pages=

    2024

  56. [68]

    Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    In prospect and retrospect: Reflective memory management for long-term personalized dialogue agents , author=. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  57. [72]

    2026 , eprint=

    OpenAI GPT-5 System Card , author=. 2026 , eprint=

    2026

  58. [73]

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, and 1 others. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  59. [74]

    Akari Asai, Zeqiu Wu, Yizhong Wang, Avi Sil, and Hannaneh Hajishirzi. 2024. Self-rag: Learning to retrieve, generate, and critique through self-reflection. In International conference on learning representations, volume 2024, pages 9112--9141

    2024

  60. [75]

    Yushi Bai, Xin Lv, Jiajie Zhang, Hongchang Lyu, Jiankai Tang, Zhidian Huang, Zhengxiao Du, Xiao Liu, Aohan Zeng, Lei Hou, and 1 others. 2024. Longbench: A bilingual, multitask benchmark for long context understanding. In Proceedings of the 62nd annual meeting of the association for computational linguistics (volume 1: Long papers), pages 3119--3137

    2024

  61. [76]

    Yushi Bai, Shangqing Tu, Jiajie Zhang, Hao Peng, Xiaozhi Wang, Xin Lv, Shulin Cao, Jiazheng Xu, Lei Hou, Yuxiao Dong, and 1 others. 2025. Longbench v2: Towards deeper understanding and reasoning on realistic long-context multitasks. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 3639--3664

    2025

  62. [77]

    Haonan Bian, Zhiyuan Yao, Sen Hu, Zishan Xu, Shaolei Zhang, Yifu Guo, Ziliang Yang, Xueran Han, Huacan Wang, and Ronghao Chen. 2026. Realmem: Benchmarking llms in real-world memory-driven interaction. arXiv preprint arXiv:2601.06966

    work page arXiv 2026

  63. [78]

    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

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  64. [79]

    Gheorghe Comanici, Eric Bieber, Mike Schaekermann, Ice Pasupat, Noveen Sachdeva, Inderjit Dhillon, Marcel Blistein, Ori Ram, Dan Zhang, Evan Rosen, and 1 others. 2025. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities. arXiv preprint arXiv:2507.06261

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  65. [80]

    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

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  66. [81]

    Shailja Gupta, Rajesh Ranjan, and Surya Narayan Singh. 2024. A comprehensive survey of retrieval-augmented generation (rag): Evolution, current landscape and future directions. arXiv preprint arXiv:2410.12837

    work page arXiv 2024

  67. [82]

    Yi Jiang, Sendong Zhao, Jianbo Li, Haochun Wang, Lizhe Zhang, Yan Liu, and Bing Qin. 2025. Cocoa: Collaborative chain-of-agents for parametric-retrieved knowledge synergy. arXiv preprint arXiv:2508.01696

    work page arXiv 2025

  68. [83]

    Ziyan Jiang, Xueguang Ma, and Wenhu Chen. 2024. Longrag: Enhancing retrieval-augmented generation with long-context llms. arXiv preprint arXiv:2406.15319

    work page arXiv 2024

  69. [84]

    Jiajie Jin, Xiaoxi Li, Guanting Dong, Yuyao Zhang, Yutao Zhu, Yongkang Wu, Zhonghua Li, Qi Ye, and Zhicheng Dou. 2025. https://arxiv.org/abs/2505.10413 Hierarchical document refinement for long-context retrieval-augmented generation . Preprint, arXiv:2505.10413

    work page arXiv 2025

  70. [85]

    Kuang-Huei Lee, Xinyun Chen, Hiroki Furuta, John Canny, and Ian Fischer. 2024. A human-inspired reading agent with gist memory of very long contexts. arXiv preprint arXiv:2402.09727

    work page arXiv 2024

  71. [86]

    u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt \

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich K \"u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt \"a schel, and 1 others. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in neural information processing systems, 33:9459--9474

    2020

  72. [87]

    Parth Sarthi, Salman Abdullah, Aditi Tuli, Shubh Khanna, Anna Goldie, and Christopher Manning. 2024. Raptor: Recursive abstractive processing for tree-organized retrieval. In International Conference on Learning Representations, volume 2024, pages 32628--32649

    2024

  73. [88]

    Aaditya Singh, Adam Fry, Adam Perelman, Adam Tart, Adi Ganesh, Ahmed El-Kishky, Aidan McLaughlin, Aiden Low, AJ Ostrow, Akhila Ananthram, Akshay Nathan, Alan Luo, Alec Helyar, Aleksander Madry, Aleksandr Efremov, Aleksandra Spyra, Alex Baker-Whitcomb, Alex Beutel, Alex Karpenko, and 467 others. 2026 a . https://arxiv.org/abs/2601.03267 Openai gpt-5 system...

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  74. [89]

    Agentic Retrieval-Augmented Generation: A Survey on Agentic RAG

    Aditi Singh, Abul Ehtesham, Saket Kumar, Tala Talaei Khoei, and Athanasios V. Vasilakos. 2026 b . https://arxiv.org/abs/2501.09136 Agentic retrieval-augmented generation: A survey on agentic rag . Preprint, arXiv:2501.09136

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  75. [90]

    Haoran Sun, Shaoning Zeng, and Bob Zhang. 2026. H-mem: Hierarchical memory for high-efficiency long-term reasoning in llm agents. In Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers), pages 341--350

    2026

  76. [91]

    Zhen Tan, Jun Yan, I-Hung Hsu, Rujun Han, Zifeng Wang, Long Le, Yiwen Song, Yanfei Chen, Hamid Palangi, George Lee, and 1 others. 2025. In prospect and retrospect: Reflective memory management for long-term personalized dialogue agents. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), page...

    2025

  77. [92]

    Yu Wang and Xi Chen. 2025. Mirix: Multi-agent memory system for llm-based agents. arXiv preprint arXiv:2507.07957

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  78. [93]

    Wujiang Xu, Zujie Liang, Kai Mei, Hang Gao, Juntao Tan, and Yongfeng Zhang. 2026. A-mem: Agentic memory for llm agents. Advances in Neural Information Processing Systems, 38:17577--17604

    2026

  79. [94]

    Sikuan Yan, Xiufeng Yang, Zuchao Huang, Ercong Nie, Zifeng Ding, Zonggen Li, Xiaowen Ma, Jinhe Bi, Kristian Kersting, Jeff Z Pan, and 1 others. 2025. Memory-r1: Enhancing large language model agents to manage and utilize memories via reinforcement learning. arXiv preprint arXiv:2508.19828

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  80. [95]

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, and 1 others. 2025. Qwen3 technical report. arXiv preprint arXiv:2505.09388

    work page internal anchor Pith review Pith/arXiv arXiv 2025

Showing first 80 references.