A team of researchers introduced a new class of language models called Context Language Models, or CLMs, which manage their own context by treating it as a file that can be updated directly by the model. This approach differs from traditional systems that rely on external mechanisms to control context length and content. By allowing the model to modify its context file without restrictions, the authors argue that CLMs can learn which information to retain and how to structure it for optimal performance. The framework extends naturally to multi-agent environments, where each agent maintains its own context file, enabling more organic coordination. Experiments show that CLMs achieve better results than existing methods while using fewer computational resources. On the BrowseComp-Plus benchmark, they delivered 11.4% higher accuracy with 21.5% fewer FLOPs. On the 12-hour EdgeBench task, they scored 5% higher using 59% fewer FLOPs. In a multi-repository agent-swarm task spanning 24 hours, the improvement reached 65% with identical compute. The authors also demonstrated that CLMs can be guided by natural-language instructions refined through optimization loops, boosting performance on specific context-management tasks by up to 35.9 points while reducing compute needs. An online reinforcement learning technique further improved Qwen3.5-9B’s results on BrowseComp-Plus by 47.6% with 12% fewer FLOPs. To enhance efficiency during deployment, the team developed Suffix Cache Reuse, a co-design strategy that cuts server-side compute by 35% compared to standard SGLang implementations at equivalent performance levels. These findings suggest that rethinking context as a mutable, model-editable structure could reshape how future language models handle information retention and reasoning over extended interactions.