Mulch Self Improver — Let your agents grow 🌱

OpenClaw skill · Cursor skill · agent skills · self-improvement skill — Uses Mulch so expertise compounds across sessions. Better and more consistent coding, improved experience, less hallucination. ClawHub-ready; works with OpenClaw, Cursor, Claude, and other AI coding agents.

---

Qualification: features, benefits & pain points

For full qualification use (sales, onboarding, fit): see QUALIFICATION.md.

Pain points we solve	Benefits	Features
Agents forget across sessions; no single project memory	Better, more consistent coding	Single learning store (.mulch/), mulch prime + mulch record
Hallucination from lack of grounding	Improved experience; less re-explaining	Typed records (failure, convention, decision, pattern, guide, reference)
Same mistakes repeated; knowledge only in chat	Less hallucination	Domains, search/query, promotion to CLAUDE/AGENTS/SOUL/TOOLS
Slow onboarding for new agents/teammates	Expertise compounds; team-wide via git	OpenClaw hook, optional scripts, mulch setup provider hooks
Scattered or ad-hoc learnings	Works with any agent; git-tracked	Robust docker-test for validation

Who it’s for: Teams using AI coding agents (Cursor, Claude, OpenClaw, Codex, etc.) who want session-to-session memory and fewer repeated errors. See QUALIFICATION.md for the full checklist.

---

How to run the Docker test

From the project root:

cd /path/to/mulch-self-improving-agent
docker build -t mulch-self-improver-test .
docker run --rm mulch-self-improver-test

Side-by-side benchmark vs Self Improving Agent — Rank #2 on ClawHub (proves Mulch is more token-efficient):

docker run --rm mulch-self-improver-test benchmark

Two nanobots run the same task series (session start, record failure + convention, retrieve “package manager” and “known failures”): one using legacy .learnings + long reminder, one using Mulch. The benchmark reports reminder chars, session context chars, retrieval chars, and asserts Mulch wins on reminder (shorter) and retrieval (targeted, so fewer chars). See BENCHMARK.md for the full table and projected savings. Troubleshooting skill improvement (quantified): Token efficiency — Mulch needs ~54% fewer chars than baseline to get the same resolutions (559 vs 1215). Find rate — baseline 3/3, Mulch ≥2/3; same or better with clear descriptions/resolutions. Interpretation — less context per run → fewer tokens, less noise, lower risk of wrong fix.

What the robust test does

1. Mulch CLI: init; add multiple domains (test, api); record all types (failure, convention, pattern, decision, reference, guide); record-then-search round-trip (record a unique convention, then mulch search and assert it’s found); mulch query / query --all; mulch prime / prime <domain>; mulch search; mulch status; mulch validate; mulch doctor.
2. OpenClaw hook (consolidated in scripts/docker-test-hook.js): Reminder token-efficient (shorter than legacy Self Improving Agent — Rank #2 on ClawHub reminder); bootstrap main → injects SELF_IMPROVEMENT_REMINDER.md; sub-agent → no injection; null/missing context / wrong type or action → no throw, no injection; bootstrapFiles missing or null → no crash; missing sessionKey → inject.
3. Scripts: activator.sh, error-detector.sh, extract-skill.sh --dry-run.
4. Skill assets: SKILL.md present with required content.

---

• Benchmark (comparison table, projected savings, troubleshooting, style & memory): BENCHMARK.md
• Benchmark in docs (baseline Rank #2 on ClawHub vs Mulch diff highlights): references/benchmark.md
• Visual benchmark breakdown (HTML): docs/benchmark-comparison.html — styled tables and KPIs (visual-explainer pattern)
• Benchmark elevator pitch (HTML): docs/benchmark-elevator-pitch.html — one-page results summary and pitch
• Qualification (features, benefits, pain points): QUALIFICATION.md
• Skill usage and workflow: SKILL.md

---

Keywords (SEO): OpenClaw skill, Cursor skill, ClawHub, agent skills, self-improvement skill, Mulch, AI coding agent, session memory, agent memory, learnings, CLAUDE.md, AGENTS.md, mulch-cli