Researchers introduce Context Language Models (CLMs), a new architecture that treats the model's context window as an editable file, allowing the model to make unrestricted updates to its own memory. This approach shifts context management from external harness control to intrinsic model behavior, enabling both in-context and parametric learning of context-management strategies.

  • Zero-shot CLMs outperform state-of-the-art context management strategies, achieving 11.4% higher accuracy with 21.5% fewer FLOPs on BrowseComp-Plus and 5% higher scores with 59% fewer FLOPs on 12-hour EdgeBench.
  • On a 24-hour multi-repository agent-swarm task, CLMs show 65% greater improvement with the same compute compared to baselines.
  • Steering CLMs with natural-language instructions evolved through a skill-optimization loop improves held-out accuracy by up to 35.9 points on a context-management task while reducing compute.
  • An online reinforcement learning method improves Qwen3.5-9B performance on BrowseComp-Plus by 47.6% while using 12% fewer FLOPs.
  • Co-designed Suffix Cache Reuse for CLM serving reduces server-side compute by 35% relative to standard SGLang at matched performance.

By allowing models to learn what is most important to maintain in context, CLMs naturally extend to multi-agent systems where multiple agent contexts coexist as files.