Context Language Models let the model edit its own memory, so it rewrites what stays in the window instead of taking whatever a summarizer or a retrieval tool left behind. The mode worth copying depends on the limit you have: zero-shot when a long job has to get cheaper with no training budget, a written skill when you want a procedure you can read, and a training run only when a smaller model has to learn the habit.
The work comes from Meta Superintelligence Labs, MIT, the University of Washington, and Trillium Labs. You can't turn it on inside Claude or ChatGPT. The research code is a lab repository you can read. What you copy from it is the file edit: the model appends a note or rewrites a stretch that's already there, and that change is in the window on the next turn. It can rewrite an old message, keep a progress note, or drop what it judges irrelevant. In a swarm, several of those files sit side by side.
Older harness tricks still fail in the usual ways. A summary can drop a detail or invent one, compaction follows rules you wrote before the run, and external memory still waits until the agent decides what to pull back. Once the model holds the file, the edit can go past those menus. One sentence in the prompt can steer when it compacts, or tell it to keep a backup, and you don't touch the weights to do that. You still choose how the habit gets in. The three modes are no extra training, a skill the run revises, and reinforcement learning on the weights (InfoQ). No extra training leaves the model you already serve, a written skill gives you a procedure you can open, and reinforcement learning costs a training run, which you take on only when a smaller model has to learn the editing habit. Don't stack them. If your agent already runs for hours, make that choice in the harness you have before you budget a training run.
What each mode is for
| Approach | Accuracy | Compute |
|---|---|---|
| Zero-shot CLM | +11.4% BrowseComp-Plus; +5% on 12-hour EdgeBench; 65% greater improvement on a 24-hour multi-repo swarm | 21.5% fewer FLOPs; 59% fewer on EdgeBench; same on the swarm |
| In-context skill evolution | Up to +35.9 points on held-out ContextBench | Lower compute than the starting skill |
| Online RL (Qwen3.5-9B) | BrowseComp-Plus 28.8% to 42.5% (47.6% relative) | 12% fewer FLOPs |
Match each row to the limit you have. Zero-shot spends no training budget. Skill evolution spends feedback on a document you can open. Online reinforcement learning spends a training run on a smaller model.
Zero-shot keeps the model you already serve and only changes how context is stored. On a model such as Qwen3.6-27B, you don't need CLM-specific training: a 12-hour repository job scores higher on a much smaller compute bill than a summary harness, and a day-long swarm that tunes several repositories together improves more at the same spend, so you can copy this without a training cluster when the job already runs for hours.
Skill evolution is the mode when you want a procedure you can read. The model keeps a document of context-management steps and revises it from feedback, while held-out ContextBench tasks grade memory hygiene apart from general reasoning, and the largest lift comes at lower compute than the skill the loop started from, which is what you want if you're tired of hand-tuning a compaction prompt.
Online reinforcement learning is the mode that changes the weights, and it's built for a smaller model. Training Qwen3.5-9B on deep-research tasks, with a reward that first requires success and then prefers the cheaper successful runs, produces the BrowseComp-Plus jump, so use it when the model you can afford barely edits on its own and skip it when you need an answer this week.
Editable context can keep a prompt injection, or an instruction the model wrote itself, alive across turns. A CLM can also throw away a fact it cannot later retrieve, and more control over the file doesn't give it better judgment about what to keep.
If you're about to rebuild an agent harness, match the mode to the limit you actually have. Use zero-shot when you can swap the context layer and a long job has to get cheaper without a training budget. Let a skill document evolve when you want a procedure you can open and argue about. Spend the reinforcement-learning run only when a smaller model has to learn the editing habit and you can afford that training. Leave the research repo in the lab until you have a plan for injections that stick and for facts an edit throws away.