{"paper":{"title":"GAAMA: Graph Augmented Associative Memory for Agents","license":"http://creativecommons.org/licenses/by-sa/4.0/","headline":"GAAMA builds a concept-mediated knowledge graph to improve memory retrieval for AI agents across multi-session conversations.","cross_cats":["cs.IR","cs.MA"],"primary_cat":"cs.AI","authors_text":"Nitin Sareen, Shubhendu Sharma, Swarna Kamal Paul","submitted_at":"2026-03-29T23:33:38Z","abstract_excerpt":"AI agents that interact with users across multiple sessions require persistent long-term memory to maintain coherent, personalized behavior. Current approaches either rely on flat retrieval-augmented generation (RAG), which loses structural relationships among memories, or use entity-centric knowledge graphs that suffer from mega-hub effects in conversational data, diluting graph-based relevance propagation. We propose GAAMA, a graph-augmented associative memory for agents that constructs a concept-mediated knowledge graph through a three-step pipeline: (1)verbatim episode preservation, (2)LLM"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"On LoCoMo-10 (1,540 questions, 10 multi-session conversations), GAAMA achieves 79.1% mean reward, a +4.2 pp improvement over a tuned RAG baseline, the strongest comparator. On MemoryArena, GAAMA outperforms full-context baselines across three tasks with advantages growing monotonically with dialogue length.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The LLM-based extraction of atomic facts and topic-level concept nodes, together with the synthesis of higher-order reflections, accurately preserves structural relationships and avoids introducing errors that would degrade retrieval quality.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"GAAMA constructs a four-node-type knowledge graph with concept-mediated edges and Personalized PageRank retrieval to improve multi-session agent memory, reporting 79.1% mean reward on LoCoMo-10 (+4.2 pp over tuned RAG) and gains on MemoryArena tasks that grow with dialogue length.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"GAAMA builds a concept-mediated knowledge graph to improve memory retrieval for AI agents across multi-session conversations.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"ed2cf21661f166bb9ef33f2e95666a544138f023be098b992ed47951423294b5"},"source":{"id":"2603.27910","kind":"arxiv","version":2},"verdict":{"id":"3ae038c2-cef8-4ef9-858e-77f11a442878","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T21:05:32.860871Z","strongest_claim":"On LoCoMo-10 (1,540 questions, 10 multi-session conversations), GAAMA achieves 79.1% mean reward, a +4.2 pp improvement over a tuned RAG baseline, the strongest comparator. On MemoryArena, GAAMA outperforms full-context baselines across three tasks with advantages growing monotonically with dialogue length.","one_line_summary":"GAAMA constructs a four-node-type knowledge graph with concept-mediated edges and Personalized PageRank retrieval to improve multi-session agent memory, reporting 79.1% mean reward on LoCoMo-10 (+4.2 pp over tuned RAG) and gains on MemoryArena tasks that grow with dialogue length.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The LLM-based extraction of atomic facts and topic-level concept nodes, together with the synthesis of higher-order reflections, accurately preserves structural relationships and avoids introducing errors that would degrade retrieval quality.","pith_extraction_headline":"GAAMA builds a concept-mediated knowledge graph to improve memory retrieval for AI agents across multi-session conversations."},"references":{"count":18,"sample":[{"doi":"","year":1990,"title":"User’s birthday is March 15, 1990","work_id":"d34d6a0c-e7d9-404f-86f2-1b2e08599e7e","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"**Do NOT duplicate existing facts.** If an existing fact already captures the information, skip it","work_id":"ac3a60e3-57cf-47e6-8902-6c8116edf459","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"**Resolve relative dates to absolute dates** using the conversation timestamp. For example, if the conversation date is \"2023-06-15\" and the user says \"last week\", resolve to approximately \"2023-06-08","work_id":"94176a33-6f0e-46d5-a8aa-5c19cd43f5ac","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Derive general knowledge from episodes by doing multi-step reasoning where possible","work_id":"bc599878-56e6-4c5f-bb0e-eee6ea174fd8","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Only extract general knowledge, preferences, attributes, or relationships that can be applied broadly","work_id":"95439836-326d-4358-af3d-bbdb7b121a72","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":18,"snapshot_sha256":"022c26f9294501f88cb1b4555bcf515c63557b3f425746e11972461a05162c1c","internal_anchors":0},"formal_canon":{"evidence_count":1,"snapshot_sha256":"8f4e7ec69ecc6e9d92bdae211d3d6299bc4727e98aae8c020c1ef1e6da55f6bc"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}