Context Language Models Let Agents Rewrite Context

CLMs make context management a model capability instead of brittle harness logic—a compelling direction for long-horizon agents, though the efficiency claims need independent validation.

• –Editable context naturally supports multi-agent workflows by giving each agent its own persistent context file.
• –Reported gains include 11.4% higher BrowseComp-Plus accuracy with 21.5% fewer FLOPs and 35% lower serving compute through Suffix Cache Reuse.
• –The approach could reduce reliance on fixed summarization, compaction, and retrieval policies.
• –Unrestricted context writes create a new prompt-injection persistence surface that production systems will need to secure.
• –The open-source repository is promising but currently offers infrastructure and experiments rather than turnkey trained checkpoints.