Cursor Composer Self-Summarization

Source: https://cursor.com/blog/self-summarization Author: Cursor team Date: 2025-2026

Summary

Cursor trains Composer to summarize its own context when hitting context limits, rather than using external summarization. Trained via RL: the reward signal applies to both problem-solving steps and summary quality. Summaries average ~1,000 tokens vs 5,000+ for manual summarization prompts, 50% fewer compaction errors.

Key Claims

  • Self-summarization baked into training — not a bolt-on external tool
  • Key insight: “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”
  • Process: generate → hit threshold → self-compress → continue with compressed context + metadata
  • Learned summaries: ~1,000 tokens (vs 5,000+ for manual prompts) + 50% error reduction
  • Demonstrated: Composer solved “make-doom-for-mips” in 170 turns, compressing 100K+ tokens
  • RL reward applies to summary quality — model learns what information is essential

Connection to Other Sources

Directly addresses the problem in Context Rot — degradation as context grows. Self-summarization is Cursor’s solution.

Entities

  • Cursor (coding agent company)

Concepts