Metacognition Skill

A self-reflection engine for AI agents. Extracts patterns from session transcripts into a weighted graph with Hebbian learning and time decay.

What It Does

• - Maintains a store of categorized insights (perceptions, overrides, protections, self-observations, decisions, curiosities)
• Uses Hebbian reinforcement: repeated insights get stronger, unused ones decay
• Builds a graph of connections between related insights
• Finds clusters of related knowledge that may represent higher-level principles
• Compiles a "metacognition lens" — a token-budgeted summary of active self-knowledge

Setup

1. 1. Place metacognition.py in your workspace scripts/ directory
2. The script stores data in memory/metacognition.json (relative to workspace)
3. The compiled lens outputs to INLINECODE3
4. Optionally configure a local embeddings endpoint for semantic similarity (falls back to string matching)

Cron Integration

Set up a cron job to run periodically (e.g., every 4 hours):

CODEBLOCK0

CLI Commands

CODEBLOCK1

Configuration

Key constants in the script:

Entry Types

• - perceptions — Things learned from experience
• overrides — Corrections to previous beliefs
• protections — Rules to prevent known failure modes
• self-observations — Patterns in own behavior
• decisions — Policy decisions for future behavior
• curiosities — Open questions with lifecycle (born → active → evolving → resolved)