Meta and UW paper lets agents edit their own context like a file: +11.4% accuracy, -21.5% compute
datawithsuman · x · 2026-10-05
A new paper from Meta Superintelligence Labs and the University of Washington (with MIT and Trillium Labs) introduces Context Language Models (CLMs): instead of a fixed script deciding what stays in context (e.g. summarize at 75% full), the model mirrors its live context as a file and edits it with plain Bash—deleting stale text, rewriting, offloading, or keeping lines verbatim—with every edit syncing back as the next turn's context.
Key results
- BrowseComp-Plus deep research: 59.4%, 11.4% above the best baseline, with 21.5% less compute
- 12-hour EdgeBench: 5% higher score, 59% less compute than Codex-style summarization
- 24-hour six-repo agent swarm: 65% greater speedup at the same compute
Memory management is learnable
- A one-line instruction changes how the model manages context
- An evolved "skill" note lifted a held-out test from 38.3% to 74.2%
- RL took Qwen3.5-9B from 28.8% to 42.5% on BrowseComp-Plus
Caveats and fixes
- Edits break KV-cache reuse (everything after the first change gets recomputed); their Suffix Cache Reuse recovers 35% server compute at matched performance
- With a 128K budget the gap to summarization mostly closes (47.3 vs 47.8), though cost stays lower
- Some math results are single runs; editable context opens a new prompt-injection surface
More from coding & agent
- New TLAPS-Bench blog: formally proving correctness of real-world system specs with TLA+ — tianyin_xu · 2026-11-19
- Pierre releases Words, a streaming Markdown parser 2x faster than md4x and 20x faster than Streamdown — andersonbcdefg · 2026-11-19
- Ex-Claude superfan: getting the best from Astra/Sol requires relearning how you prompt — mariofilhoml · 2026-11-19
- Rivet's lesson: JEV doesn't reason—split severity classification into tiny yes/no prompts — yawnxyz · 2026-11-19
- 16x Microsoft MVP demos a native PowerShell agent built on Microsoft's Intelligent Terminal — dfinke · 2026-11-19
- Musk: Treat Grok Bot like a really smart employee, says most users underuse it — elonmusk · 2026-11-19