Cursor Composer 2's compaction-in-loop RL cuts mid-task 'forgot the constraint' errors ~50%

Cursor·medium signal

Composer 2 (released March 19, 2026) trains the model to compress its own ~200K context down to roughly 1,000 tokens mid-session and continue — and because the compression happens inside the RL loop with a reward for maintaining task completion post-compression, the model learns which details to keep, reporting ~50% fewer compaction errors than external summarization. The practitioner signal: 'compaction-in-the-loop' is becoming the differentiator for multi-hour autonomous coding, and external bolt-on summarizers are the weak point. Builders evaluating coding agents for long refactors should test specifically for constraint loss after compression, not just raw context size.