Meta and UW paper lets agents edit their own context as a file

Researchers from UW and Meta propose Context Language Models, which manage their own context via Bash edits and beat compaction baselines.

Researchers from the University of Washington and Meta Superintelligence Labs, with MIT and Trillium Labs, posted "Context Language Models" on 29 September 2026. The idea is to treat an agent's context as a file the model can freely edit with Bash. The model decides what to keep, rewrite or delete, instead of following harness rules for compaction, offloading or retrieval. If you run long agent loops, this is a technique you can test today with models you already call. It is a research result, not a shipped product feature.

The authors say CLMs built zero-shot outperform the context-management baselines they tested: 11.4% higher accuracy with 21.5% fewer FLOPs on BrowseComp-Plus, and 5% higher scores with 59% fewer FLOPs on 12-hour EdgeBench. The full paper names Qwen3.6-27B and GPT5.6-Sol among the zero-shot models. It also reports an online RL method that improves Qwen3.5-9B on BrowseComp-Plus by 47.6%. Mid-context edits break prefix caching, so the team added Suffix Cache Reuse, which they say cuts server-side compute by 35% against standard SGLang.

The code includes a minimal CLM agent for Harbor tasks and a Pi package. It is licensed CC BY-NC 4.0, which rules out commercial use. A DEV Community write-up notes three more points. No trained weights were released. The context file is discarded when a task ends. The paper's safety section warns that this freedom lets a model plant instructions for itself.

Some questions remain unclear from the sources. They don't show what stops a model from deleting critical instructions, since edits are "unrestricted." They also don't show whether the code plugs into harnesses such as Claude Code.

Sources