Learnings Skill

Continuous learning from failures, corrections, and patterns. Logs mistakes to prevent repetition. Own MeiliSearch index. Learns version rules, approach preferences, and behavioral corrections.

Continuous learning system. Every failure, correction, and pattern gets logged so the same mistakes aren't repeated. Uses its own MeiliSearch index for fast recall.

Security Model

• Credentials: MeiliSearch key is loaded from ~/.openclaw/workspace/.env — never hardcoded in scripts
• Sensitive data filtering: Auto-extraction skips content containing tokens, passwords, API keys, and credentials
• Safe defaults: Distillation and auto-extract default to --dry-run; use --apply to write
• Secure temp files: All temp files use mktemp with chmod 600 and are cleaned up on exit

Capabilities (declared)

• Read: memory/*.md (daily notes, for auto-extraction)
• Write: MeiliSearch learnings index, LEARNINGS.md (distillation only with --apply)
• Execute: curl to localhost MeiliSearch, python3 for document processing

What Gets Logged

Category	Description	Example
failure	Script/command errors	"npm install fails with permission error"
correction	Fix applied after a mistake	"Use --prefix flag for global installs"
approach	Right way to do something	"Use full-text search over vector DB for local memory"
version_rule	How versions should increment	"1.1.0 → 1.2.0 for new features"
preference	User preferences discovered	"Prefers local/no-cloud solutions"
recurring_failure	Same thing failing repeatedly	"Auth token expires on every deploy"
rule	Explicit "don't do X" rules	"Don't commit secrets to repositories"
learning	General lessons learned	"Always check file exists before copying"

Configuration

Requires ~/.openclaw/workspace/.env:

bash

MEILI_HOST=http://127.0.0.1:7700
MEILI_KEY=your-master-key-here

Files

• LEARNINGS.md — Human-readable summary (auto-generated, use --apply to write)
• scripts/setup_index.sh — Create/configure MeiliSearch index
• scripts/log_learning.sh — Log a new learning entry
• scripts/search_learnings.sh — Search past learnings before acting
• scripts/auto_extract.sh — Auto-extract learnings from daily notes
• scripts/distill_learnings.sh — Summarize raw logs into LEARNINGS.md

Usage

bash

# Log a learning
bash scripts/log_learning.sh \
  --category failure \
  --what "deployment fails with auth token expiry" \
  --fix "Refresh token before each deploy" \
  --importance 0.9 \
  --tags "deploy,auth,token"

# Search before acting
bash scripts/search_learnings.sh "github push" 5

# Auto-extract from daily notes (dry-run by default)
bash scripts/auto_extract.sh --days 7 --apply

# Distill into readable summary (dry-run by default)
bash scripts/distill_learnings.sh --apply

How It Integrates

Before taking any action, I search the learnings index:

1. Search for the action/context
2. If a past learning exists, apply the fix first
3. If the action fails, log it immediately
4. If you correct me, log it as a high-priority correction

Version Rules (Learned)

• Increment minor (1.1.0 → 1.2.0) for new features
• Increment major (1.0.0 → 2.0.0) for breaking changes
• Avoid patch-level bumps (1.1.0 → 1.1.1) unless it's an actual bug fix