{"paper":{"title":"MEM1: Learning to Synergize Memory and Reasoning for Efficient Long-Horizon Agents","license":"http://creativecommons.org/licenses/by/4.0/","headline":"MEM1 trains agents to keep constant memory in long multi-turn tasks by updating one shared state that merges memory and reasoning via reinforcement learning.","cross_cats":["cs.AI","cs.IR"],"primary_cat":"cs.CL","authors_text":"Alok Prakash, Ao Qu, Bryan Kian Hsiang Low, Daniela Rus, Jinhua Zhao, Paul Pu Liang, Sunghwan Kim, Zhaoxuan Wu, Zijian Zhou","submitted_at":"2025-06-18T19:44:46Z","abstract_excerpt":"Modern language agents must operate over long-horizon, multi-turn interactions, where they retrieve external information, adapt to observations, and answer interdependent queries. Yet, most LLM systems rely on full-context prompting, appending all past turns regardless of their relevance. This leads to unbounded memory growth, increased computational costs, and degraded reasoning performance on out-of-distribution input lengths. We introduce MEM1, an end-to-end reinforcement learning framework that enables agents to operate with constant memory across long multi-turn tasks. At each turn, MEM1 "},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"MEM1-7B improves performance by 3.5x while reducing memory usage by 3.7x compared to Qwen2.5-14B-Instruct on a 16-objective multi-hop QA task, and generalizes beyond the training horizon.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That reinforcement learning on composed multi-turn environments will produce a memory-update policy that reliably retains all information needed for future interdependent queries while discarding only truly irrelevant content.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"MEM1 uses end-to-end RL to learn constant-memory agents that update a shared state for memory and reasoning, delivering 3.5x better performance and 3.7x lower memory use than larger baselines on long-horizon QA and shopping tasks.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"MEM1 trains agents to keep constant memory in long multi-turn tasks by updating one shared state that merges memory and reasoning via reinforcement learning.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"ead35034d430dd1311e76ea6ecbff20594593adf1dcf49f9694339ed010db178"},"source":{"id":"2506.15841","kind":"arxiv","version":2},"verdict":{"id":"8e4d137c-11e8-489a-9a21-fd366fc6660b","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T00:22:57.994140Z","strongest_claim":"MEM1-7B improves performance by 3.5x while reducing memory usage by 3.7x compared to Qwen2.5-14B-Instruct on a 16-objective multi-hop QA task, and generalizes beyond the training horizon.","one_line_summary":"MEM1 uses end-to-end RL to learn constant-memory agents that update a shared state for memory and reasoning, delivering 3.5x better performance and 3.7x lower memory use than larger baselines on long-horizon QA and shopping tasks.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That reinforcement learning on composed multi-turn environments will produce a memory-update policy that reliably retains all information needed for future interdependent queries while discarding only truly irrelevant content.","pith_extraction_headline":"MEM1 trains agents to keep constant memory in long multi-turn tasks by updating one shared state that merges memory and reasoning via reinforcement learning."},"references":{"count":75,"sample":[{"doi":"","year":2024,"title":"Surprising exercises that will sharpen your short- term memory, January 2024","work_id":"dfa227ab-174a-4b6a-8511-4b756b3da24a","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"Why does the effective context length of llms fall short? In Proceedings of the International Conference on Learning Representations (ICLR), 2025","work_id":"5e744371-ff26-48ad-abd7-f26bbe3559c7","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"The claude 3 model family: Opus, sonnet, haiku","work_id":"fa4d3ce0-da16-49ff-bf18-86ab716b01b1","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":1974,"title":"Baddeley and Graham J","work_id":"f6fc23b6-b5aa-4982-9fff-08da2c170dac","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"Digirl: Training in-the-wild device-control agents with autonomous reinforcement learning","work_id":"a1622f76-1559-4413-bbd3-5de78aa0af9a","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":75,"snapshot_sha256":"ed8a7fa3c5fe3c55c308fd0e3b26e4355b3dbd711ad2925ed9d945ae1106ece0","internal_anchors":19},"formal_canon":{"evidence_count":3,"snapshot_sha256":"ead7ac152b6ecfee3957a83df8244804bbb7da2af6d43009c612d184b18c5a1d"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}