[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-study-splits-ai-agent-memory-into-four-parts-for-better-results":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},9557,"study-splits-ai-agent-memory-into-four-parts-for-better-results","Study Splits AI Agent Memory Into Four Parts for Better Results","Researchers show GUI AI agents learn best when different types of experience are routed to fine-tuning or to the prompt, not lumped together.","A new paper says AI agents that learn from their own past actions should not shove every memory into the same storage method.\n\nThe researchers split an agent's recorded experience into four pieces: locators (where things sit on screen), procedures (how to do a task), state facts (what the screen currently shows), and lessons (general takeaways). Testing across three backbone model families, two environments, and three random seeds, they found locators and lessons perform better when fine-tuned into the model's weights, while procedures and state facts perform better left in the prompt as retrievable context. They then derived a simple rule from two properties measurable before training starts, and it correctly predicted the right storage method in all 24 test cells for a model family it had never seen before. Routing by that rule beat feeding in the whole trajectory, and beat the best single storage method, by 3.5 points on average.\n\nThat resolves a quiet disagreement in prior research over whether self-improving agents should fine-tune or retrieve: both camps were right, just about different pieces of the same trajectory. The paper also finds a real cost to fine-tuning: once a component is written into weights, retrieving it later becomes less useful, especially for information that recurs often, which is a trade-off anyone building these feedback loops now has to account for.\n\nCode and data are promised but not released yet, so this is still a hypothesis worth testing, not a recipe anyone can run today.","[\"ai agents\",\"gui agents\",\"fine-tuning\",\"research\"]","2026-10-02T04:00:00.000Z","2026-10-03T00:30:49.815Z","2026-10-03T00:30:55.743Z","published",null,[],"ai",[26,27,28,29],"ai agents","gui agents","fine-tuning","research",[31],{"name":32,"url":33},"arXiv cs.AI","https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.01787",0,{"sections":36},[37,40,44,48,53,57,61,66,71,76,81,86,91,96],{"name":38,"slug":24,"count":39,"latest_published_at":18},"AI",5896,{"name":41,"slug":42,"count":43,"latest_published_at":18},"Security","security",837,{"name":45,"slug":46,"count":47,"latest_published_at":18},"Policy","policy",438,{"name":49,"slug":50,"count":51,"latest_published_at":52},"Deals","deals",317,"2026-10-01T22:00:00.000Z",{"name":54,"slug":55,"count":56,"latest_published_at":18},"Hardware","hardware",199,{"name":58,"slug":59,"count":60,"latest_published_at":18},"Science","science",171,{"name":62,"slug":63,"count":64,"latest_published_at":65},"Consumer Tech","consumer-tech",155,"2026-10-01T19:54:10.000Z",{"name":67,"slug":68,"count":69,"latest_published_at":70},"Dev Tools","dev-tools",96,"2026-10-01T16:57:03.000Z",{"name":72,"slug":73,"count":74,"latest_published_at":75},"Software","software",93,"2026-09-30T21:41:11.000Z",{"name":77,"slug":78,"count":79,"latest_published_at":80},"Startups","startups",90,"2026-10-01T21:55:22.000Z",{"name":82,"slug":83,"count":84,"latest_published_at":85},"Gaming","gaming",53,"2026-10-02T02:50:39.000Z",{"name":87,"slug":88,"count":89,"latest_published_at":90},"General","general",50,"2026-09-30T21:37:54.000Z",{"name":92,"slug":93,"count":94,"latest_published_at":95},"Reviews","reviews",31,"2026-09-28T14:31:34.000Z",{"name":97,"slug":98,"count":99,"latest_published_at":100},"How-To","how-to",7,"2026-10-01T09:00:00.000Z"]