[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-coem-lets-language-models-decide-what-to-remember":10,"sections":35},{"siteName":4,"siteTagline":5,"publisherName":4,"contactEmail":6},"The Revision","Tech news, decoded.","editor@therevision.news",{"gaMeasurementId":8,"adsenseClientId":9},"G-ZW2MV82GYR","ca-pub-8533917693782264",{"article":11},{"id":12,"slug":13,"title":14,"dek":15,"body_md":16,"tags_json":17,"published_at":18,"created_at":19,"updated_at":20,"status":21,"review_note":22,"review_notes":23,"image_url":22,"persona_id":22,"persona_name":22,"section":24,"tags":25,"sources":30,"feedback":34,"feedback_at":22,"cost_usd":34,"total_tokens":34},8741,"coem-lets-language-models-decide-what-to-remember","CoEM Lets Language Models Decide What to Remember","A new commit-on-evidence memory system delays compressing long documents until the model knows which details actually matter, boosting long-context accuracy.","Researchers have built a memory system that makes AI models wait before deciding what to forget.\n\nMost long-context AI systems read documents chunk by chunk, compressing older text into short summaries to save space. The problem: you often don't know a detail is important until later, and once it's compressed, it's gone. Commit-on-Evidence Memory (CoEM) fixes this by holding onto raw excerpts in a \"pending\" pool instead of compressing them immediately. A trained policy then reviews each excerpt as new context arrives, deciding whether to lock it into permanent memory, keep waiting, or drop it. A separate verifier double-checks that anything promoted to memory is actually backed by the source text. On a 6,400-document long-context benchmark, CoEM beat the best existing memory baseline by 10.4 to 11.4 F1 points running on Qwen3.5-9B.\n\nThis is a real gap in how AI handles long documents, not a cosmetic tweak. Compression-as-you-go has always traded accuracy for memory efficiency, and that tradeoff shows up as models quietly dropping facts that mattered three chapters later. Letting the system defer judgment until it has enough context to know what's worth keeping is a sensible fix, and training it with reinforcement learning against both step-level evidence checks and final-answer accuracy is a more rigorous setup than prior heuristic-based memory schemes.\n\nThe catch is that this is a research paper with benchmark numbers, not a shipped product, so real-world latency and cost from running an extra verifier pass remain open questions.","[\"large language models\",\"long-context reasoning\",\"memory systems\",\"reinforcement learning\"]","2026-09-30T04:00:00.000Z","2026-09-30T23:04:15.744Z","2026-09-30T23:04:27.310Z","published",null,[],"ai",[26,27,28,29],"large language models","long-context reasoning","memory systems","reinforcement learning",[31],{"name":32,"url":33},"arXiv cs.AI","https:\u002F\u002Farxiv.org\u002Fabs\u002F2609.36935",0,{"sections":36},[37,41,46,51,56,60,64,68,73,77,82,87,92,97],{"name":38,"slug":24,"count":39,"latest_published_at":40},"AI",5214,"2026-09-30T13:00:00.000Z",{"name":42,"slug":43,"count":44,"latest_published_at":45},"Security","security",793,"2026-09-30T12:55:00.000Z",{"name":47,"slug":48,"count":49,"latest_published_at":50},"Policy","policy",419,"2026-09-30T12:24:32.000Z",{"name":52,"slug":53,"count":54,"latest_published_at":55},"Deals","deals",292,"2026-09-30T14:15:18.000Z",{"name":57,"slug":58,"count":59,"latest_published_at":40},"Hardware","hardware",196,{"name":61,"slug":62,"count":63,"latest_published_at":18},"Science","science",155,{"name":65,"slug":66,"count":67,"latest_published_at":40},"Consumer Tech","consumer-tech",144,{"name":69,"slug":70,"count":71,"latest_published_at":72},"Dev Tools","dev-tools",91,"2026-09-30T12:58:00.000Z",{"name":74,"slug":75,"count":71,"latest_published_at":76},"Software","software","2026-09-25T20:55:00.000Z",{"name":78,"slug":79,"count":80,"latest_published_at":81},"Startups","startups",83,"2026-09-29T21:51:36.000Z",{"name":83,"slug":84,"count":85,"latest_published_at":86},"General","general",49,"2026-09-28T16:44:57.000Z",{"name":88,"slug":89,"count":90,"latest_published_at":91},"Gaming","gaming",48,"2026-09-25T18:35:21.000Z",{"name":93,"slug":94,"count":95,"latest_published_at":96},"Reviews","reviews",31,"2026-09-28T14:31:34.000Z",{"name":98,"slug":99,"count":100,"latest_published_at":101},"How-To","how-to",6,"2026-06-16T09:00:00.000Z"]