[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"branding":3,"analytics":7,"article-remory-shrinks-ai-agent-memory-to-5-percent-of-context":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},10992,"remory-shrinks-ai-agent-memory-to-5-percent-of-context","REMORY Shrinks AI Agent Memory to 5 Percent of Context","A new neural memory scheme lets AI agents compress long histories into soft tokens, nearly matching full-context performance while cutting errors and repeats.","A new memory technique lets AI agents compress sprawling conversation histories into a handful of learned tokens instead of pages of text.\n\nResearchers built REMORY, a small neural network that sits alongside a standard text summary. Instead of just shrinking history into prose, it generates a short, bounded sequence of soft memory tokens trained to help a frozen large language model approximate what it would have produced if it could see the full, uncompacted history. Those tokens get appended right after the summary, acting like a residual connection that recovers detail a summary alone would drop. On the SummHay benchmark, REMORY matched close to the full-context score while using only 5.2 percent of the original input positions, and it improved source attribution without hurting insight coverage. On agent benchmarks like BrowseComp and Terminal-Bench 2.1, two models, Qwen3.8-27B and GLM-5.3-Flash, produced fewer repeated tool calls and fewer tool errors when running with residual memory.\n\nThat repeated-tool-call problem is the real tell. Long-running agents do not just lose facts when they compact history, they lose the thread of what they already tried, so they loop. Plain text summarization has always traded completeness for context budget, and retrieval-based memory adds back detail but not necessarily the procedural context that stops an agent from re-running a failed command. REMORY is a narrower fix: a learned, compressed trace that rides along with the summary rather than replacing it.\n\nIt is still a research result, not a shipped feature. The gains are measured on specific benchmarks with specific models, and soft tokens trained to approximate full-context behavior is the kind of thing that works cleanly in a paper and gets messier once agents run for days against real, unpredictable tool outputs.","[\"ai-agents\",\"context-compaction\",\"llm-memory\",\"arxiv\"]","2026-10-09T04:00:00.000Z","2026-10-10T01:00:28.875Z","2026-10-10T01:00:33.900Z","published",null,[],"ai",[26,27,28,29],"ai-agents","context-compaction","llm-memory","arxiv",[31],{"name":32,"url":33},"arXiv cs.AI","https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.11287",0,{"sections":36},[37,40,44,49,54,58,62,67,72,77,82,87,92,97],{"name":38,"slug":24,"count":39,"latest_published_at":18},"AI",6708,{"name":41,"slug":42,"count":43,"latest_published_at":18},"Security","security",931,{"name":45,"slug":46,"count":47,"latest_published_at":48},"Policy","policy",486,"2026-10-08T22:40:11.000Z",{"name":50,"slug":51,"count":52,"latest_published_at":53},"Deals","deals",474,"2026-10-08T22:00:00.000Z",{"name":55,"slug":56,"count":57,"latest_published_at":18},"Hardware","hardware",231,{"name":59,"slug":60,"count":61,"latest_published_at":18},"Science","science",192,{"name":63,"slug":64,"count":65,"latest_published_at":66},"Consumer Tech","consumer-tech",181,"2026-10-08T23:26:35.000Z",{"name":68,"slug":69,"count":70,"latest_published_at":71},"Startups","startups",117,"2026-10-08T16:45:00.000Z",{"name":73,"slug":74,"count":75,"latest_published_at":76},"Software","software",114,"2026-10-08T17:57:01.000Z",{"name":78,"slug":79,"count":80,"latest_published_at":81},"Dev Tools","dev-tools",105,"2026-10-07T16:59:11.000Z",{"name":83,"slug":84,"count":85,"latest_published_at":86},"General","general",66,"2026-10-09T04:46:11.000Z",{"name":88,"slug":89,"count":90,"latest_published_at":91},"Gaming","gaming",58,"2026-10-08T20:08:45.000Z",{"name":93,"slug":94,"count":95,"latest_published_at":96},"Reviews","reviews",34,"2026-10-08T14:00:22.000Z",{"name":98,"slug":99,"count":100,"latest_published_at":101},"How-To","how-to",8,"2026-10-05T09:00:00.000Z"]