metacognition
meimakes 作
AI & LLMs概要
Self-reflection engine for AI agents.
説明
このスキルでできること
Extracts patterns from AI agent session transcripts into a weighted knowledge graph. Repeated insights strengthen over time through Hebbian reinforcement, while unused ones decay. Compiles a token-budgeted summary called the metacognition lens for use in future sessions.
Persists cross-session learning with weighted decay, avoiding the loss of accumulated context that comes with static system prompts or per-session memory alone.
使いどころ
- Identifying recurring mistakes an agent makes across sessions
- Tracking user communication preferences over time
- Automatically pruning outdated or weakened beliefs via decay
- Discovering higher-level principles from specific session events
- Running scheduled self-reflection cycles on a cron timer
ワークフロー例
User asks: run the metacognition integration cycle
- 1
Run
python3 scripts/metacognition.py decayto prune weak entries - 2
Read recent session transcripts using sessions_list and sessions_history
- 3
Analyze conversation for recurring patterns, relationship signals, and confidence mismatches
- 4
Add 1-3 principle-level insights with typed categories such as overrides, decisions, or curiosities
- 5
Run reweave then compile to rebuild graph connections and regenerate the metacognition lens
Updated metacognition-lens.md with reinforced insights and newly extracted principles, ready for injection into the next session context
インストール
clawhub install meimakes/metacognitionCLIコマンドを実行するか、Claude Code、Cursor、Codexなどお使いのコーディングエージェントにプロンプトを貼り付けてください。