Context Language Models let AI agents edit their own working memory

Researchers from the University of Washington, Meta Superintelligence Labs, MIT and Trillium Labs introduced Context Language Models (CLMs), which give a model its live context as an editable file. On BrowseComp-Plus, CLM reached 59.4% accuracy versus 53.4% for the strongest baseline while using 21.5% fewer prefix-reuse FLOPs.

• CLM lets the model add, rewrite, remove and reorganize its own context
• BrowseComp-Plus: 59.4% accuracy vs 53.4% for Codex-style summarization
• 21.5% fewer prefix-reuse FLOPs on the same benchmark
• On a 12-hour EdgeBench subset CLM scored ~5% higher with 59% fewer FLOPs