What is this?
This is the official code release for 'Context Language Models (CLMs)', research from University of Washington, Meta Superintelligence Labs, MIT and Trillium Labs. CLMs are language models that natively manage their own context by treating it as a file they can freely edit, which naturally extends to multi-agent setups where each agent's context is its own file.
This is the official code release for 'Context Language Models (CLMs)', research from University of Washington, Meta Superintelligence Labs, MIT and Trillium Labs. CLMs are language models that natively manage their own context by treating it as a file they can freely edit, which naturally extends to multi-agent setups where each agent's context is its own file.Is this a finished tool or research code for reproducing the paper's experiments?It's research code accompanying the CLM paper, not a polished product: the repo ships Harbor-based implementations (clmharness, clmicl, clmrl) plus a suffix-cache-reuse serving component, meant for running and reproducing the paper's experiments.
It's research code accompanying the CLM paper, not a polished product: the repo ships Harbor-based implementations (clmharness, clmicl, clmrl) plus a suffix-cache-reuse serving component, meant for running and reproducing the paper's experiments.How do I run a CLM?Run the minimal CLM agent on any Harbor task using the clm-harbor CLI with the '-a clm-minimal' flag; configuration and serving details are documented in clm/clmharness.
Run the minimal CLM agent on any Harbor task using the clm-harbor CLI with the '-a clm-minimal' flag; configuration and serving details are documented in clm/clmharness.Do I need the Harbor framework to use this?Yes, CLMs are built on top of Harbor and are invoked through the clm-harbor CLI rather than run standalone.
Yes, CLMs are built on top of Harbor and are invoked through the clm-harbor CLI rather than run standalone.Is everything from the paper already available here?Not entirely: ContextBench, the benchmark referenced for evaluating CLMs, is still listed as 'coming soon' and isn't included in the repo yet.
Not entirely: ContextBench, the benchmark referenced for evaluating CLMs, is still listed as 'coming soon' and isn't included in the repo yet.Can I use this commercially?No, the project is released under CC BY-NC 4.0, which allows research and non-commercial use only.
No, the project is released under CC BY-NC 4.0, which allows research and non-commercial use only.How much better are CLMs than standard context-management approaches?In zero-shot tests, CLMs reach 11.4% higher accuracy with 21.5% fewer FLOPs on BrowseComp-Plus, 5% higher scores with 59% fewer FLOPs on 12-hour EdgeBench, and a 65% greater improvement at equal compute on a 24-hour multi-repository agent-swarm task.
In zero-shot tests, CLMs reach 11.4% higher accuracy with 21.5% fewer FLOPs on BrowseComp-Plus, 5% higher scores with 59% fewer FLOPs on 12-hour EdgeBench, and a 65% greater improvement at equal compute on a 24-hour multi-repository agent-swarm task.Can CLM behavior be tuned without retraining the model?Yes, clm/clmicl shows CLMs can be steered with natural-language instructions evolved through a skill-optimization loop, improving held-out accuracy by up to 35.9 points on a context-management task while using less compute.
Yes, clm/clmicl shows CLMs can be steered with natural-language instructions evolved through a skill-optimization loop, improving held-out accuracy by up to 35.9 points on a context-management task while using less compute.Does the repo include a reinforcement learning method for CLMs?Yes, clm/clmrl implements an online RL method that improved Qwen3.5-9B's performance on BrowseComp-Plus by 47.6% while using 12% fewer FLOPs.
Yes, clm/clmrl implements an online RL method that improved Qwen3.5-9B's performance on BrowseComp-Plus by 47.6% while using 12% fewer FLOPs.What is suffixcachereuse for?It's a component for efficiently serving CLMs by reusing suffix caches, aimed at reducing the compute cost of running CLMs in production-like settings.
It's a component for efficiently serving CLMs by reusing suffix caches, aimed at reducing the compute cost of running CLMs in production-like settings.What are similar projects?MSA, Awesome-Graph-Engineering, Raven, awesome-rsi, Tiersense, GVS5H, jeeves, Attention-Residuals
MSA, Awesome-Graph-Engineering, Raven, awesome-rsi, Tiersense, GVS5H, jeeves, Attention-ResidualsWhich topics is it in?arxiv, foundation-models, long-term-memory, multi-agent-systems, llm-agents, paper, training, natural-language-processing, context-engineering, academic, harness, agent-framework
arxiv, foundation-models, long-term-memory, multi-agent-systems, llm-agents, paper, training, natural-language-processing, context-engineering, academic, harness, agent-framework