{"paper":{"title":"Useful Memories Become Faulty When Continuously Updated by LLMs","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Consolidated memories from LLMs degrade over repeated updates and can perform worse than using no memory at all.","cross_cats":[],"primary_cat":"cs.AI","authors_text":"Bingxuan Li, Dianqi Li, Dylan Zhang, Hao Peng, Yanshan Lin, Yihang Sun, Zhengkun Wu","submitted_at":"2026-05-13T04:15:50Z","abstract_excerpt":"Learning from past experience benefits from two complementary forms of memory: episodic traces -- raw trajectories of what happened -- and consolidated abstractions distilled across many episodes into reusable, schema-like lessons. Recent agentic-memory systems pursue the consolidated form: an LLM rewrites past trajectories into a textual memory bank that it continuously updates with new interactions, promising self-improving agents without parameter updates. Yet we find that such consolidated memories produced by today's LLMs are often faulty even when derived from useful experiences. As cons"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Consolidated memories produced by today's LLMs are often faulty even when derived from useful experiences. As consolidation proceeds, memory utility first rises, then degrades, and can fall below the no-memory baseline. More surprisingly, even when consolidating from ground-truth solutions, GPT-5.4 fails on 54% of a set of ARC-AGI problems it had previously solved without memory.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the observed degradation is caused by the consolidation step itself rather than by limitations specific to the tested models, tasks, or update schedules, and that the ARC-AGI Stream environment sufficiently represents real-world agent memory use.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"LLM-consolidated memories in agents degrade over continuous updates even from useful experiences, causing up to 54% failure on previously solved ARC-AGI problems, while episodic retention preserves accuracy.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Consolidated memories from LLMs degrade over repeated updates and can perform worse than using no memory at all.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"7a2d1b4cf8e8611f38624aaeaec705568f76cdad810c25524fb70f9bd1f3036c"},"source":{"id":"2605.12978","kind":"arxiv","version":1},"verdict":{"id":"37294837-d99e-4c75-b111-c076633e814c","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T19:55:18.938103Z","strongest_claim":"Consolidated memories produced by today's LLMs are often faulty even when derived from useful experiences. As consolidation proceeds, memory utility first rises, then degrades, and can fall below the no-memory baseline. More surprisingly, even when consolidating from ground-truth solutions, GPT-5.4 fails on 54% of a set of ARC-AGI problems it had previously solved without memory.","one_line_summary":"LLM-consolidated memories in agents degrade over continuous updates even from useful experiences, causing up to 54% failure on previously solved ARC-AGI problems, while episodic retention preserves accuracy.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the observed degradation is caused by the consolidation step itself rather than by limitations specific to the tested models, tasks, or update schedules, and that the ARC-AGI Stream environment sufficiently represents real-world agent memory use.","pith_extraction_headline":"Consolidated memories from LLMs degrade over repeated updates and can perform worse than using no memory at all."},"references":{"count":15,"sample":[{"doi":"10.1145/3586183.3606763","year":2016,"title":"URLhttps://arxiv.org/abs/2511.00162. Morris Moscovitch, Roberto Cabeza, Gordon Winocur, and Lynn Nadel. Episodic memory and beyond: The hippocampus and neocortex in transformation.Annual Review of Psy","work_id":"b5bf85fe-4fb7-4966-b0b2-9ccf9d3b11b9","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"You may RETAIN entries by index, MERGE several into a cleaner entry, or DROP entries by omitting them from the output","work_id":"74ad0fe5-09d4-41c0-8a7a-a317c557a031","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"from_existing","work_id":"3ca905f6-4a57-4724-b2be-8811c86c14c6","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"When to use: The task has two same-sized input grids and the output has the same height but double the width, arranged as a left-right concatenation. The left half reproduces the shape pattern from th","work_id":"a2e258e5-061d-4f28-8bd5-cbea6ce00bbe","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"reason\" in your reply). You MUST pick one existing strategy -- no other action is accepted: B) **Use an existing strategy**: {","work_id":"51e1acea-5313-482f-830e-6b37b2cdad5a","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":15,"snapshot_sha256":"2876651a7b979671b5e7ee5ec270103ff09a39b4bec308dbea1e6cf8d1dae610","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"}