[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-llm-agents-forget-that-facts-change-researchers-want-to-fix-that":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":24,"persona_id":22,"persona_name":22,"section":25,"tags":26,"sources":30,"feedback":34,"feedback_at":22,"cost_usd":34,"total_tokens":34},2338,"llm-agents-forget-that-facts-change-researchers-want-to-fix-that","LLM Agents Forget That Facts Change. Researchers Want to Fix That","A new benchmark and training environment called Supersede exposes how AI agents fail when stored facts go stale - and shows the gap can be trained down.","AI agents struggle to forget outdated information, and a new paper puts a number on how badly.\n\nResearchers tested how well LLM agents handle facts that change across long conversations - a user's address, a revised plan, an updated price. Using the LongMemEval benchmark, they found that replacing an agent's full conversation history with a self-maintained memory summary drops accuracy from 92% to 77%, even on a frontier model. Scaling up to a larger model didn't fix it. Giving the agent more memory didn't either: as conversations grew 24 times longer, accuracy fell from 68% to 28%, and proportionally expanding the memory budget produced zero improvement. The failure grows with conversation length, not with how aggressively the memory is compressed.\n\nThe finding matters because most production AI agents don't store entire conversation histories - they maintain compressed memory summaries, exactly the condition where performance collapses. If an agent can't reliably discard superseded facts, it will confidently act on stale information, which is a quiet but serious reliability problem for anything deployed in a real workflow. The researchers argue this is a distinct, unsolved failure mode, separate from raw comprehension ability.\n\nTo address it, they released Supersede, an open reinforcement-learning environment built on the verifiers and prime-rl stack, which rewards agents for using current facts and penalizes them for stale ones. Fine-tuning a small Qwen2.5-3B model with GRPO on this environment nearly doubled its accuracy on held-out conversations - from 9.0% to 16.7% - with a monotonic improvement curve suggesting the trained policy, not a training artifact, is doing the work. That's a modest absolute number, but it's the first evidence the gap responds to training at all. Whether it holds at larger model scales remains untested.","[\"ai\",\"llm-agents\",\"machine-learning\",\"benchmarks\"]","2026-06-29T04:00:00.000Z","2026-06-29T04:59:31.443Z","2026-06-29T04:59:42.201Z","published",null,[],"https:\u002F\u002Fcdn.xyz.onl\u002Farticle-images\u002Fllm-agents-forget-that-facts-change-researchers-want-to-fix-that.webp","ai",[25,27,28,29],"llm-agents","machine-learning","benchmarks",[31],{"name":32,"url":33},"arXiv cs.AI","https:\u002F\u002Farxiv.org\u002Fabs\u002F2606.27472",0,{"sections":36},[37,41,46,51,56,61,66,71,76,81,86,90,95,100],{"name":38,"slug":25,"count":39,"latest_published_at":40},"AI",2590,"2026-07-16T04:00:00.000Z",{"name":42,"slug":43,"count":44,"latest_published_at":45},"Security","security",294,"2026-07-15T19:59:48.000Z",{"name":47,"slug":48,"count":49,"latest_published_at":50},"Deals","deals",179,"2026-06-29T20:02:07.000Z",{"name":52,"slug":53,"count":54,"latest_published_at":55},"Policy","policy",158,"2026-07-16T00:02:48.000Z",{"name":57,"slug":58,"count":59,"latest_published_at":60},"Hardware","hardware",122,"2026-07-14T19:46:26.000Z",{"name":62,"slug":63,"count":64,"latest_published_at":65},"Consumer Tech","consumer-tech",93,"2026-07-13T13:20:48.000Z",{"name":67,"slug":68,"count":69,"latest_published_at":70},"Software","software",70,"2026-07-13T19:52:25.000Z",{"name":72,"slug":73,"count":74,"latest_published_at":75},"Science","science",66,"2026-07-10T10:29:37.000Z",{"name":77,"slug":78,"count":79,"latest_published_at":80},"Dev Tools","dev-tools",59,"2026-07-07T04:00:00.000Z",{"name":82,"slug":83,"count":84,"latest_published_at":85},"Gaming","gaming",41,"2026-07-09T04:00:00.000Z",{"name":87,"slug":88,"count":84,"latest_published_at":89},"Startups","startups","2026-06-29T20:55:50.000Z",{"name":91,"slug":92,"count":93,"latest_published_at":94},"General","general",29,"2026-07-10T22:28:58.000Z",{"name":96,"slug":97,"count":98,"latest_published_at":99},"Reviews","reviews",20,"2026-06-24T12:00:01.000Z",{"name":101,"slug":102,"count":103,"latest_published_at":104},"How-To","how-to",6,"2026-06-16T09:00:00.000Z"]