By making self-summarization part of Composer's training, we can get training signal from trajectories much longer than the model's max context window.

Cursor

Cursor's Composer model is trained to handle long-horizon coding tasks via a technique called self-summarization, integrated directly into its reinforcement learning training loop. When Composer approaches its context limit, it pauses to generate a condensed summary of its own context before continuing. Because this summarization is part of training, the model learns to retain only the most critical information. Compared to a heavily engineered prompt-based compaction baseline, Composer's self-summarization reduces compaction error by 50% while using one-fifth of the tokens. As a case study, an early Composer 2 checkpoint solved a challenging Terminal-Bench 2.0 problem (compiling Doom for MIPS) over 170 turns, self-summarizing 100,000+ tokens down to ~1,000 tokens multiple times. The team sees this as a stepping stone toward multi-agent coordination and even longer task horizons.

Training Composer for longer horizons · Cursor

# Self-summarization as a trained behavior

<p>the self-summarization as a trained behavior is the interesting part here. most approaches to context compression are post-hoc engineering hacks. what cursor is doing is baking the compaction behavior into the RL loop itself so the model learns when and how to summarize its own context as part of the task completion strategy.</p>
<p>the implication for long horizon agentic tasks is significant. composer can now plan and execute over sessions that would have previously hit hard context walls. the paper on token-efficient compaction is worth reading in full if you’re building anything that involves agents working across large codebases. the difference between a model that runs out of context and one that self-summarizes in-flight is the difference between needing human intervention and fully autonomous execution</p>
