ClawHub Skill Spotlight: self-improving-agent — Your AI Gets Smarter With Every Use

🟡 Installation Verified

What Problem Does This Skill Solve?

Ever run into this? AI makes a mistake today, makes the exact same mistake tomorrow. You correct it three times and it still does it a fourth.

The AI hasn't gotten dumber — it simply has no cross-session memory. When a conversation ends, every lesson learned, every pitfall encountered, is wiped clean.

self-improving-agent was built to fix exactly this.

Skill Overview

Item	Details
Skill Name	self-improving-agent
Author	@pskoett
ClawHub Downloads	263,000+ (site-wide #1)
Stars	⭐ 2,400+
Current Version	v3.0.5 (17 historical versions)
License	MIT-0 (free to use, no attribution required)
Security Scan	VirusTotal ✅ Benign / OpenClaw ✅ Benign high confidence

What Does It Actually Do?

Once installed, the AI Agent automatically logs information to the .learnings/ directory when it encounters any of the following:

Trigger	Logged To
Command execution fails	`.learnings/ERRORS.md`
User corrects the AI	`.learnings/LEARNINGS.md` (category: correction)
User requests a feature	`.learnings/FEATURE_REQUESTS.md`
API call fails	`.learnings/ERRORS.md` (with integration details)
Knowledge gap identified	`.learnings/LEARNINGS.md` (category: knowledge_gap)
Better approach found	`.learnings/LEARNINGS.md` (category: best_practice)

These logs aren't just for show — when an issue is generic enough, it can be promoted to global memory files:

SOUL.md → AI personality and behavioral guidelines
AGENTS.md → workflow and automation rules
TOOLS.md → tool usage tips and pitfalls

In essence, this system gives the AI an "experience journal" — letting it accumulate knowledge across sessions.

Installation

Method 1: ClawHub CLI (Recommended)

clawhub install self-improving-agent

Method 2: Manual Install

git clone https://github.com/peterskoett/self-improving-agent.git \
  ~/.openclaw/skills/self-improving-agent

Post-Install Initialization

mkdir -p ~/.openclaw/workspace/.learnings

Then create three log files:

.learnings/LEARNINGS.md — experience and correction logs
.learnings/ERRORS.md — error logs
.learnings/FEATURE_REQUESTS.md — feature request logs

Optional: Enable Hooks (Auto-Reminders)

cp -r hooks/openclaw ~/.openclaw/hooks/self-improvement
openclaw hooks enable self-improvement

Once enabled, the AI receives a "Should I log any learnings?" reminder at the start of each session.

Real-World Test Results

✅ ClawHub install flow is smooth — done with one command
✅ Directory structure is clear — the three files in .learnings/ each serve a distinct purpose
✅ Standardized log format — each entry includes ID, timestamp, priority, status, and suggested fix
✅ Sensible promotion threshold — issues must appear 3+ times before being promoted to global rules, keeping noise out
✅ Compatible with multiple Agents: OpenClaw, Claude Code, Codex, GitHub Copilot all supported

Things to Watch Out For

Hooks remind, they don't auto-record: Enabling hooks only injects a reminder at session start. Whether the AI actually logs anything depends on the model. Sonnet/Opus executes more consistently than Haiku.
Token overhead: Each session adds roughly 50–100 tokens of context overhead — lightweight but present.
Permissions note: The skill injects content into your workspace and writes files. OpenClaw's security scan recommends "checking the install target before enabling."
Sub-Agent compatibility: v3.0.5 fixes the sub-agent session failure bug.

Recommendation

⭐⭐⭐⭐⭐ 5/5 — Highly Recommended

Best for:

Heavy OpenClaw users with multiple long sessions daily
Teams using AI for continuous iterative development on complex projects
Anyone who has made the same mistake more than twice and wants to do something about it

Not ideal for:

Casual AI users with occasional, light-use sessions
Users who want minimal context overhead (hooks add a small number of tokens)

Bottom line:

This is ClawHub's most downloaded skill for a reason. If you use OpenClaw every day, this should be the first skill you install.

SFD Lab | Tested: 2026-03 | OpenClaw v1.x