---
title: CLM (Context Language Models)
type: vocabulary
url: "https://www.envisioning.com/vocab/clm-context-language-models"
summary: Language models that manage their own context as an editable file rather than an append-only transcript, letting them decide what to keep and extending naturally to multi-agent systems.
year: 2026
generality: 0.45
---

# CLM (Context Language Models)

Language models that manage their own context as an editable file rather than an append-only transcript, letting them decide what to keep and extending naturally to multi-agent systems.
CLM (Context Language Models) describes language models that manage their own context directly, instead of relying on an external harness to decide what stays in the conversation log. Rulin Shao, Nathan Lambert, Luke Zettlemoyer, Pang Wei Koh, and ten coauthors at Meta Superintelligence Labs, the University of Washington, MIT, and Trillium Labs introduced the term in "Context Language Models," posted to arXiv (2609.37725) on 29 September 2026. It is a preprint, not yet peer reviewed. The acronym CLM already names a different, unrelated Envisioning entry, Contrastive Language Model, a dual-encoder action scorer. The two share only the initials.

The mechanism: the model treats its context as a file it can edit without restriction, rather than an append-only transcript that only grows. This lets the model learn what is worth keeping and what to discard, and it extends naturally to multi-agent systems, where each agent's context is a separate file. Built zero-shot from existing models, the authors report CLMs beating standard, harness-controlled context management: 11.4% higher accuracy with 21.5% fewer FLOPs on BrowseComp-Plus, 5% higher scores with 59% fewer FLOPs on a twelve-hour EdgeBench task, and 65% greater improvement at matched compute on a twenty-four-hour multi-repository agent-swarm task.

Because context handling moves from the harness into the model, the paper shows two ways to steer it further. Natural-language instructions evolved through a skill-optimization loop raise held-out accuracy on a context-management task by up to 35.9 points while cutting compute. An online reinforcement-learning method lifts Qwen3.5-9B's BrowseComp-Plus score by 47.6% using 12% fewer FLOPs. The same paper introduces Suffix Cache Reuse, a serving-side KV-cache technique built for CLMs that cuts server-side compute by 35% against standard SGLang at matched performance, published as code at github.com/facebookresearch/context-language-models.

---
Source: Envisioning — Technology Research Institute (https://www.envisioning.com/vocab/clm-context-language-models)