Learn to manage your own context
Today the harness manages a model’s context with fixed rules, like summarizing every N steps. A new MIT paper on Context Language Models lets the model manage its own context instead: the context is a file the model can edit with Bash. On deep-research tasks this gave 11.4% higher accuracy with 21.5% less compute.
Takeaways
- Given control, models invent their own habits: a private notes role, live progress ledgers updated in place, and small scripts that trim old search results.
- Editing the middle of the context breaks the usual KV cache. The paper reuses the cache after the edit and trains the model to edit surgically.
- Context management is moving from the harness into the model. Any such method has to explain its edits, or users lose track of what the model was told.
Have a system that needs a second opinion?