Context Language Models (Meta)

Meta Superintelligence Labs + UW's open-source harness for 'Context Language Models': models that natively manage their own context by treating it as an editable file, with zero-shot, in-context-learning and RL recipes plus Suffix Cache Reuse, a KV-cache optimization patched into SGLang.

Why it matters

What you could build with it

Does it hold up?

Too early to judge independently — only the paper's own numbers exist, no third-party replication yet, and the code ships under CC BY-NC 4.0, so it is research/evaluation-grade rather than a commercial drop-in; the pi-clm extension shows day-one community pickup.

Built with Context Language Models

• pi-clm: the agent that manages its own contextgithub · Pi agent extension implementing the CLM loop — the model edits a context-mirror file each turn, and a /clm panel charts context size with compaction points and per-revision diffs. 19 stars, MIT.
• Context Language Models: what the UW and Meta paper changes for agent buildersdev.to · Builder-focused analysis arguing CLMs absorb the hand-engineered context-management layer into the model, and that providers will likely fold it into existing model families rather than ship a separate product line.
• UW-Meta paper: language models that edit their own contextaiweekly · News alert summarizing the BrowseComp-Plus and EdgeBench claims, with the caveat that all numbers come from the paper itself and there is no third-party replication yet.
• lolipopshock/pi-clmgithub · 19 ★ · Pi agent extension that implements the CLM context-as-a-file loop with a compaction panel; installable via 'pi install npm:@lolipopshock/pi-clm'.

Learn more

video explainer →

First spotted on github: source.

More AI infrastructure