{"paper":{"title":"Cooperative Memory Paging with Keyword Bookmarks for Long-Horizon LLM Conversations","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Cooperative paging with keyword bookmarks lets LLMs recover evicted conversation turns on demand and achieve higher answer quality than full context or retrieval methods.","cross_cats":["cs.AI"],"primary_cat":"cs.CL","authors_text":"Ziyang Liu","submitted_at":"2026-04-14T07:06:35Z","abstract_excerpt":"When LLM conversations grow beyond the context window, old content must be evicted -- but how does the model recover it when needed? We propose cooperative paging: evicted segments are replaced with minimal keyword bookmarks ([pN:keywords], ~8-24 tokens each), and the model is given a recall() tool to retrieve full content on demand. On the LoCoMo benchmark (10 real multi-session conversations, 300+ turns), cooperative paging achieves the highest answer quality among six methods -- outperforming truncation, BM25, word-overlap retrieval, a search-tool baseline, and full context -- on four model"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"On the LoCoMo benchmark (10 real multi-session conversations, 300+ turns), cooperative paging achieves the highest answer quality among six methods -- outperforming truncation, BM25, word-overlap retrieval, a search-tool baseline, and full context -- on four models (GPT-4o-mini, DeepSeek-v3.2, Claude Haiku, GLM-5), confirmed by four independent LLM judges (p=0.017, paired bootstrap).","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That keyword bookmarks supply enough distinctive information for the model to correctly trigger and select the right page via the recall() tool, given the reported 57% correct selection rate when bookmarks lack specificity.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Cooperative paging replaces evicted LLM context with keyword bookmarks and adds a recall tool, outperforming six other methods on the LoCoMo benchmark across four models with statistical significance.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Cooperative paging with keyword bookmarks lets LLMs recover evicted conversation turns on demand and achieve higher answer quality than full context or retrieval methods.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"479463706bf92a7e932eec911d8ad1253953ea1607bfa88100b67ac76263762d"},"source":{"id":"2604.12376","kind":"arxiv","version":2},"verdict":{"id":"57840bdd-8026-4dee-9606-963b51535503","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-10T15:35:26.791767Z","strongest_claim":"On the LoCoMo benchmark (10 real multi-session conversations, 300+ turns), cooperative paging achieves the highest answer quality among six methods -- outperforming truncation, BM25, word-overlap retrieval, a search-tool baseline, and full context -- on four models (GPT-4o-mini, DeepSeek-v3.2, Claude Haiku, GLM-5), confirmed by four independent LLM judges (p=0.017, paired bootstrap).","one_line_summary":"Cooperative paging replaces evicted LLM context with keyword bookmarks and adds a recall tool, outperforming six other methods on the LoCoMo benchmark across four models with statistical significance.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That keyword bookmarks supply enough distinctive information for the model to correctly trigger and select the right page via the recall() tool, given the reported 57% correct selection rate when bookmarks lack specificity.","pith_extraction_headline":"Cooperative paging with keyword bookmarks lets LLMs recover evicted conversation turns on demand and achieve higher answer quality than full context or retrieval methods."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2604.12376/integrity.json","findings":[],"available":true,"detectors_run":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}