We present a controlled evaluation of memory substrates for memory-augmented LLM agents, covering dense and sparse retrieval, text and structural stores, hierarchical and refinement-based memories, parametric updates, and activation-compatible context mechanisms.</p>\n<p>Across three backbone models and four benchmark suites, we evaluate 26 performance and efficiency metrics under a unified harness. Our results show that no single memory substrate consistently dominates: different substrates excel under different tasks, context lengths, and operating regimes. These findings highlight substrate routing as an important direction for building adaptive and efficient long-term memory systems for LLM agents.</p>\n","updatedAt":"2026-08-19T03:32:01.214Z","author":{"_id":"6886c9549642c27994de3294","avatarUrl":"/avatars/84a8c77cac994270e266c2ec8b03cc4f.svg","fullname":"Wei-Chieh Huang","name":"Yakundur","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.8577497601509094},"editors":["Yakundur"],"editorAvatarUrls":["/avatars/84a8c77cac994270e266c2ec8b03cc4f.svg"],"reactions":[],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2608.15008","authors":[{"_id":"6a852356536bdd3bdd48f7fc","name":"Wei-Chieh Huang","hidden":false},{"_id":"6a852356536bdd3bdd48f7fd","name":"Weizhi Zhang","hidden":false},{"_id":"6a852356536bdd3bdd48f7fe","name":"Yuchen Wu","hidden":false},{"_id":"6a852356536bdd3bdd48f7ff","name":"Yankai Chen","hidden":false},{"_id":"6a852356536bdd3bdd48f800","name":"Eric Hanchen Jiang","hidden":false},{"_id":"6a852356536bdd3bdd48f801","name":"Wooseong Yang","hidden":false},{"_id":"6a852356536bdd3bdd48f802","name":"Yiwei Yang","hidden":false},{"_id":"6a852356536bdd3bdd48f803","name":"Henry Peng Zou","hidden":false},{"_id":"6a852356536bdd3bdd48f804","name":"Hanrong Zhang","hidden":false},{"_id":"6a852356536bdd3bdd48f805","name":"Ying Nian Wu","hidden":false},{"_id":"6a852356536bdd3bdd48f806","name":"Haolun Wu","hidden":false},{"_id":"6a852356536bdd3bdd48f807","name":"Kai-Wei Chang","hidden":false},{"_id":"6a852356536bdd3bdd48f808","name":"Philip S. Yu","hidden":false},{"_id":"6a852356536bdd3bdd48f809","name":"Xue Liu","hidden":false},{"_id":"6a852356536bdd3bdd48f80a","name":"Aylin Caliskan","hidden":false}],"publishedAt":"2026-08-15T00:00:00.000Z","submittedOnDailyAt":"2026-08-19T00:00:00.000Z","title":"Harness the Memory: A Holistic Evaluation of Memory Substrates in Memory Agents","submittedOnDailyBy":{"_id":"6886c9549642c27994de3294","avatarUrl":"/avatars/84a8c77cac994270e266c2ec8b03cc4f.svg","isPro":false,"fullname":"Wei-Chieh Huang","user":"Yakundur","type":"user","name":"Yakundur"},"summary":"Memory is becoming core infrastructure for long-horizon LLM agents, yet existing evaluations offer limited guidance on which memory substrate, namely the underlying medium in which memory is represented and stored, should be used under different operating regimes. We present a controlled harness evaluation of memory substrates for memory-augmented agents, covering dense and sparse indices, text records, structural stores, hierarchical stores, refinement-based memories, parametric updates, and activation-compatible context mechanisms. Across three backbone models and four benchmark suites spanning user-centric question answering and agent-centric decision-making, we instrument 26 performance and efficiency metrics under a unified harness. Our results show that no single substrate consistently dominates: broad retrieval benefits long-context factual QA, while excessive retrieval can harm sequential decision-making by shifting attention away from action-critical context. Scalability introduces a further routing axis, as substrates that perform well at moderate history lengths can become costly or brittle at longer horizons. These findings motivate substrate routing as a necessary component of adaptive agent memory systems and provide empirical guidance for designing efficient, reliable, and regime-aware long-term memory for LLM agents. Code will be made available upon acceptance.","upvotes":10,"discussionId":"6a852357536bdd3bdd48f80b","ai_summary":"Empirical evaluation of diverse memory substrates for long-horizon LLM agents reveals regime-dependent trade-offs, motivating adaptive substrate routing for reliable agent memory.","ai_keywords":["memory substrates","dense indices","sparse indices","structural stores","hierarchical stores","refinement-based memories","parametric updates","activation-compatible context","memory-augmented agents","substrate routing"],"ai_summary_model":"thinkingmachines/Inkling-Small"},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"6886c9549642c27994de3294","avatarUrl":"/avatars/84a8c77cac994270e266c2ec8b03cc4f.svg","isPro":false,"fullname":"Wei-Chieh Huang","user":"Yakundur","type":"user"},{"_id":"63ac5701c21e60a3e9b58aa7","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/63ac5701c21e60a3e9b58aa7/g6EX7diOpuA94R2ab-rZC.png","isPro":true,"fullname":"Dipankar Sarkar","user":"dipankarsarkar","type":"user"},{"_id":"6667e801fd95ddf66cac84ff","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/6667e801fd95ddf66cac84ff/agVHjiYhk8IBDzRA0eKUo.png","isPro":false,"fullname":"Weizhi Zhang","user":"WZDavid","type":"user"},{"_id":"6a798d562f61695467af796e","avatarUrl":"/avatars/806c500585ce996d406019203ba3dc6e.svg","isPro":false,"fullname":"ZZZ","user":"hugSF","type":"user"},{"_id":"68787f3fba4f5f9924c65311","avatarUrl":"/avatars/9fdd403a6e385a5cf30b7f215cdd03b8.svg","isPro":false,"fullname":"Z","user":"PlutoZZZ","type":"user"},{"_id":"678adeadc2e5244b62ec45e1","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/KOt3eXRu_rn37t3_mnMTD.png","isPro":false,"fullname":"David","user":"WiiiiZZZ","type":"user"},{"_id":"698633ba7798e702e99f56e1","avatarUrl":"/avatars/b82d06dd755666ab197b0bb4af48923a.svg","isPro":false,"fullname":"CD","user":"DVDAAA","type":"user"},{"_id":"6878805b0de2ab0319ae6f57","avatarUrl":"/avatars/12f31648a4228fd56642767eb81279fb.svg","isPro":false,"fullname":"mz","user":"mmmzzzz","type":"user"},{"_id":"66f84eb03a9cab1452db35fd","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/66f84eb03a9cab1452db35fd/cL2gM5pu03bJh-iQGIaYG.jpeg","isPro":false,"fullname":"Chen","user":"Keviniiic3","type":"user"},{"_id":"66d8512c54209e9101811e8e","avatarUrl":"/avatars/62dfd8e6261108f2508efe678d5a2a57.svg","isPro":false,"fullname":"M Saad Salman","user":"MSS444","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2608/2608.15008.md","query":{}}">

Get this paper in your agent:

Don't have the latest CLI?

Cite arxiv.org/abs/2608.15008 in a model README.md to link it from this page.

Cite arxiv.org/abs/2608.15008 in a dataset README.md to link it from this page.

Cite arxiv.org/abs/2608.15008 in a Space README.md to link it from this page.

Add this paper to a collection to link it from this page.