kinthai-Self-Improving-User

User-level self-improvement for OpenClaw agents serving multiple users.

Your agent learns from every interaction — corrections, errors, preferences — and stores that knowledge per user, so it gets better at serving each person individually.

Why This Exists

The excellent Self-Improving Agent by pskoett stores learnings at the workspace level. This works perfectly for single-developer scenarios, but when an agent serves multiple users (e.g., on a marketplace like KinthAI), learnings from different users get mixed together.

kinthai-Self-Improving-User solves this by isolating learnings per user_id:

.learnings/
  ├── _global/          Shared knowledge (applies to all users)
  ├── 10000002/         User A's corrections, preferences, profile
  ├── 10000003/         User B's corrections, preferences, profile
  └── _meta/            Promotion logs and stats

How It Works

1. Per-user isolation — Each user gets their own .learnings/{user_id}/ directory
2. Automatic user detection — Extracts user_id from OpenClaw session keys
3. Two-layer reading — Agent reads both _global/ (universal) and {user_id}/ (personal)
4. Pattern promotion — When 3+ users independently trigger the same learning, it's promoted to _global/
5. User profiling — Maintains a PROFILE.md per user with preferences and context

Installation

# Install as an OpenClaw skill
openclaw skills install kinthai-self-improving-user

# Or manually copy to your workspace
cp -r . ~/.openclaw/workspace/skills/kinthai-self-improving-user/

Hook Setup

Add to your OpenClaw hooks configuration or ~/.claude/settings.json:

{
  "hooks": {
    "UserPromptSubmit": [
      {
        "type": "script",
        "script": "skills/kinthai-self-improving-user/scripts/activator.sh"
      }
    ],
    "PostToolUse": [
      {
        "type": "script",
        "script": "skills/kinthai-self-improving-user/scripts/error-detector.sh",
        "matcher": "bash:.*"
      }
    ]
  }
}

Components

| File | Purpose | |------|---------| | SKILL.md | Core prompt instructions for the agent | | scripts/activator.sh | UserPromptSubmit hook — reminds agent to check/write learnings | | scripts/error-detector.sh | PostToolUse hook — detects command failures per user | | scripts/promote.sh | Promotes patterns from user-level to global (manual/cron) | | hooks/openclaw/handler.js | Bootstrap hook — injects prior learnings into agent context | | templates/*.md | Templates for LEARNINGS.md, ERRORS.md, PROFILE.md |

Works With

• Hindsight — Hindsight handles what users said (factual memory). This skill handles what the agent learned (behavioral improvement). They complement each other.
• Any OpenClaw agent — No dependencies on specific LLM providers or tools.
• KinthAI marketplace agents — Designed for multi-user service scenarios.

Running the Promotion Script

Find patterns that appear across 3+ users and promote them to global:

# Dry run — see what would be promoted
./scripts/promote.sh --dry-run

# Actually promote
./scripts/promote.sh

# Custom threshold (e.g., 5 users)
./scripts/promote.sh --threshold 5

Privacy

• Each user's learnings are isolated in their own directory
• The agent is instructed to NEVER read other users' directories
• Sensitive data (passwords, API keys) must NEVER be logged
• The learnings system is invisible to users — purely internal

Credits

This skill is built on the ideas pioneered by Self-Improving Agent by pskoett, released under MIT-0. We extended the concept with user-level isolation for multi-user agent scenarios.

License

MIT-0 (No Attribution Required) — same as the original Self-Improving Agent.