What This Skill Does
One line: It makes AI automatically learn from its mistakes and user corrections, writing lessons to files so the same error never happens twice.
Once installed, the skill triggers automatically when:

User says "that's wrong" / "you forgot again" / "this doesn't work"
A command fails with an error
User requests a capability that doesn't exist yet
The AI realizes its knowledge is outdated or incorrect
A better approach is discovered

These experiences get categorized into:

.learnings/LEARNINGS.md — behavioral improvement records
.learnings/ERRORS.md — error and failure logs
.learnings/FEATURE_REQUESTS.md — unmet capability requests

Installation and Configuration

Install the skill

clawhub install self-improving-agent

The skill is located at:

~/.openclaw/workspace/skills/self-improving-agent/SKILL.md
No additional configuration needed. The skill is auto-discovered through OpenClaw's available_skills system and activates when matching scenarios occur.
Real-World Usage — SFD Lab Scenarios
Scenario 1: API Path Mistake Auto-Recorded
When integrating the ssls.vip domain registration API, we used /order/create and kept getting 404s. The correct endpoint was /order/buy.
With self-improving-agent, this correction was automatically logged to .learnings/ERRORS.md:

2026-03-15 | SSLS.vip API Wrong Path

Error: Using /order/create caused 404
Fix: POST /order/buy (not /order/create)
Note: Must include template_id=130 or registration has no template
Source: Direct correction from boss
Next time this API is needed, the AI checks learnings first. Same mistake: never again.
Scenario 2: Command Failure Automatically Learned
When generating cover images on the server, our initial code used require('canvas'). The server didn't have this package installed — immediate failure.
The failure was logged:

2026-03-27 | Node.js canvas Module Missing

Context: Server-side cover image generation
Error: Cannot find module 'canvas'
Solution: Server only has sharp. Use SVG→WebP pipeline instead of canvas drawing
Path: /var/www/cms-oss/node_modules/sharp
Future similar tasks: go straight to sharp, skip canvas.
Scenario 3: Review Learnings Before Major Tasks
The skill description says "Also review learnings before major tasks." In SFD Lab, before every deployment run, we have the AI review the learnings file to confirm there are no known pitfalls lurking.
Gotchas and Lessons
Gotcha 1: The Skill Records, Doesn't Auto-Retrieve
Self-improving-agent is a recording tool, not a search tool. It writes experiences to files, but it won't automatically scan all historical errors before every response unless you explicitly trigger it.
Solution: Add a rule to SOUL.md or AGENTS.md: "Before major tasks, read .learnings/LEARNINGS.md." This ensures it runs at session start.
Gotcha 2: Files Grow Indefinitely
If every error gets appended, the file gets huge over months, slowing down reads.
Solution: Periodic "distillation" — merge duplicate entries, archive resolved issues to .learnings/archive/.
Gotcha 3: Vague Entries Are Useless
Writing "API call failed" without specific parameters, error messages, or solutions is worthless when you look it up later.
Solution: Standard format for ERRORS.md: Date + Scenario + Specific Error + Root Cause + Solution + Source. All fields required.
Combining with Other Skills
With self-evolution
self-evolution is the higher-level system that can modify core files like SOUL.md. self-improving-agent is the foundation layer, focused on error recording. Together: use self-improving-agent daily, run self-evolution for major version upgrades periodically.
With coding-agent
After every ACP coding task that hits errors or requires rework, auto-trigger self-improving-agent to log the fix. Coding solutions accumulate in learnings instead of vanishing in conversation history.
With skill-self-evolution-enhancer
More powerful combination: gives any skill its own dedicated evolution capability. For example, give 12306-train-assistant its own learnings specifically for 12306 API quirks and changes.
How SFD Lab Uses This Skill
In our 13-agent collaboration system, self-improving-agent plays the role of "organizational memory":

🔥 Dragon (CEO): Reviews learnings after each task to confirm experiences were captured
🐝 Bee (Deployment): Logs deployment failures immediately to projects/[project]/.learnings/
🦅 Falcon (Security): Security audit findings and fixes go to a security-specific learnings file
🐙 Octopus / 🐺 Wolf (Backend/Fullstack): Code errors and solutions logged in real-time

Key design: each project has its own .learnings/ directory, not a single global file. Project-specific experiences stay organized and don't bleed into each other.
Summary
Self-improving-agent solves the most common problem in AI systems: making the same mistake twice.
Near-zero installation cost. Benefits compound over time. If you're running a long-term project or multi-agent system, this skill should be in your baseline configuration.
SFD Lab Note: We've used this skill for 21 days at SFD Lab. The learnings files now contain 37 valid entries. Most direct value: that "domain registration must include template_id=130" trap — we stepped in it exactly once.