promote-lessons

Promote Lessons

Converts episodic memory (corrections, lessons) into procedural memory (behavioral rules). The script queries mem for correction-tagged cells and high-salience lesson cells, synthesizes terse behavioral rules via LLM, and injects them between HTML markers in your agent's config file.

Works with OpenClaw (SOUL.md), Claude Code (CLAUDE.md), and any agent platform that reads a project file at startup.

How It Works

1. mem search-tag correction → finds cells tagged with correction
2. mem scene lessons → finds high-salience cells (sal ≥ 0.8) in the lessons scene
3. Combined cells → LLM prompt → numbered list of terse behavioral rules
4. Rules injected between <!-- LEARNED_RULES_START --> / <!-- LEARNED_RULES_END --> in target file

Target File Auto-Detection

The script picks the target in this order:

1. Explicit TARGET_FILE env var
2. SOUL.md in $WORKSPACE (OpenClaw convention)
3. CLAUDE.md in $WORKSPACE (Claude Code convention)

API Provider Auto-Detection

The script auto-detects the LLM provider:

• If API_URL contains anthropic.com → uses Anthropic Messages API (x-api-key auth)
• Otherwise → uses OpenAI Chat Completions API (Bearer auth, works with OpenClaw)

Override with API_PROVIDER=anthropic or API_PROVIDER=openclaw.

Setup

1. Add HTML markers to your target file

Add this block where you want learned rules to appear:

<!-- LEARNED_RULES_START -->
## Learned Rules
*No rules yet — run promote-lessons after tagging some corrections.*
<!-- LEARNED_RULES_END -->

2. Configure and test

OpenClaw:

export API_TOKEN=$(python3 -c "import json; print(json.load(open('$HOME/.openclaw/openclaw.json'))['gateway']['auth']['token'])")
export WORKSPACE="/path/to/your/workspace"
bash promote-lessons.sh

Claude Code (Anthropic API):

export API_TOKEN="sk-ant-your-key-here"
export API_URL="https://api.anthropic.com/v1/messages"
export WORKSPACE="/path/to/your/project"
bash promote-lessons.sh