Same tasks, fewer tokens.

Mulch Self Improver uses ~27.5% fewer characters than the legacy Self Improving Agent (Rank #2 on ClawHub) for session start + troubleshooting + style & memory lookups.

27.5%

Total efficiency gain (3792 vs 5233 chars; ~352 tokens saved)

Session: Shorter reminder + mulch prime → ~14% fewer chars.
Troubleshooting: Targeted mulch search instead of full ERRORS.md + LEARNINGS.md → ~54% fewer chars.
Style & memory: Domain separation + targeted search → ~33% fewer chars.

Run: docker run --rm mulch-self-improver-test benchmark · Full tables: benchmark-comparison.html