Self Improving Agent vs Mulch

Baseline (Rank #2 on ClawHub, legacy .learnings) vs Mulch Self Improver — what’s different

Self Improving Agent — Rank #2 on ClawHub

Store	`.learnings/` (LEARNINGS.md, ERRORS.md, mixed PREFERENCES-style file)
Session start	Long reminder (632 chars) + load full .learnings files into context
Recording	Append to markdown files (no types, no domains)
Retrieval	Read full file(s); grep/cat to find “package manager” or errors
Troubleshooting	Load full ERRORS.md + LEARNINGS.md (1215 chars) to find a fix
Style / memory	One mixed file (e.g. PREFERENCES.md); load full file for any question

Mulch Self Improver

Store	`.mulch/` (typed records, domains)
Session start	Short reminder (452 chars) + `mulch prime`; only prime output in context
Recording	`mulch record <domain> --type failure\|convention\|…`
Retrieval	`mulch search "…"` / `mulch query` (targeted; 330 chars vs 932 for 2 queries)
Troubleshooting	One `mulch search "<error>"` per scenario (559 chars, ~54% less)
Style / memory	Domains `writing_style`, `addressing`, `preferences`, `habits`, `admin`; targeted search (757 vs 1136 chars, ~33% less)

Net: Mulch uses a shorter reminder, targeted retrieval (search/query) instead of full-file reads, and domain separation for errors vs conventions vs style/preferences, so the same tasks use fewer characters (token proxy) and less noise. Combined efficiency gain over legacy: ~27.5% fewer chars when session + troubleshooting + style/memory are all used (3792 vs 5233 chars; ~352 tokens saved).

Full benchmark: benchmark-comparison.html · Elevator pitch: benchmark-elevator-pitch.html