{"paper":{"title":"M$^\\star$: Every Task Deserves Its Own Memory Harness","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"LLM agents perform better with task-specific memory programs evolved as Python code than with any fixed shared design.","cross_cats":["cs.AI","cs.CL","cs.LG"],"primary_cat":"cs.PL","authors_text":"Mirror Xu, Shiwei Zhang, Shujie Liu, Wanlu Shi, Wenbo Pan, Xiangyang Zhou, Xiaohua Jia","submitted_at":"2026-04-10T03:22:26Z","abstract_excerpt":"Large language model agents rely on specialized memory systems to accumulate and reuse knowledge during extended interactions. Recent architectures typically adopt a fixed memory design tailored to specific domains, such as semantic retrieval for conversations or skills reused for coding. However, a memory system optimized for one purpose frequently fails to transfer to others. To address this limitation, we introduce M$^\\star$, a method that automatically discovers task-optimized memory harnesses through executable program evolution. Specifically, M$^\\star$ models an agent memory system as a "},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"M$^* improves performance over existing fixed-memory baselines robustly across all evaluated tasks. Furthermore, the evolved memory programs exhibit structurally distinct processing mechanisms for each domain.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That reflective population-based code evolution on the chosen benchmarks will reliably discover memory programs that are both superior and generalizable beyond the specific evaluation tasks and failure modes analyzed.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"M* evolves distinct Python memory programs per task via population-based reflective search, outperforming fixed-memory baselines on conversation, planning, and reasoning benchmarks.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"LLM agents perform better with task-specific memory programs evolved as Python code than with any fixed shared design.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"feac4ba2381661eef6410860175d7320b30380b7196a390f4c7caeb7e37d229e"},"source":{"id":"2604.11811","kind":"arxiv","version":2},"verdict":{"id":"492efbc2-7f1c-4924-aa07-ace7911b35ea","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-10T17:34:40.321222Z","strongest_claim":"M$^* improves performance over existing fixed-memory baselines robustly across all evaluated tasks. Furthermore, the evolved memory programs exhibit structurally distinct processing mechanisms for each domain.","one_line_summary":"M* evolves distinct Python memory programs per task via population-based reflective search, outperforming fixed-memory baselines on conversation, planning, and reasoning benchmarks.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That reflective population-based code evolution on the chosen benchmarks will reliably discover memory programs that are both superior and generalizable beyond the specific evaluation tasks and failure modes analyzed.","pith_extraction_headline":"LLM agents perform better with task-specific memory programs evolved as Python code than with any fixed shared design."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2604.11811/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"}