Meta-Learning Engine

The Problem

Your agent makes the same mistakes across sessions. It learns facts but not lessons. It accumulates information but not wisdom. Every correction you make evaporates at the next context reset.

What This Does

Three-phase meta-learning loop that runs end-of-session or via cron:

Phase 1: Extract — Scans sessions for learning signals: failures, corrections, surprises, decisions, pushback. Not "what happened" but "what was learned."

Phase 2: Distill — Compresses signals into three output types:

• Regressions — "Don't do X again" (auto-appends to AGENTS.md)
• Principles — "When Y happens, do Z" (proposes SOUL.md updates, requires human approval)
• Skill Updates — Concrete improvements to skill files

Phase 3: Compound — Tracks learning velocity over time. Which domains are improving? Where is the agent plateauing? Weekly growth reports with plateau detection.

The Philosophy

Built on principles from Josh Waitzkin (The Art of Learning), Derek Sivers, and James Carse:

• Failure is Material — Every mistake becomes a guardrail. Zero wasted failures.
• Making Smaller Circles — Tracks repeated topics and goes deeper, not wider.
• Incremental over Entity — Measures growth rate, not fixed capability.
• Obvious to You, Amazing to Others — Surfaces insights the agent would discard.

How It Changes Your Agent

Week 1: Captures 5-10 regressions. Stops repeating mistakes. Week 2-3: Patterns emerge. First principle proposals. Rejections are learning signals too. Month 1: Personalized regression list, 5-10 earned principles, measurable domain improvement. Month 3+: Plateau detection triggers depth over breadth. The agent is measurably different.

What You Get

• meta-learn.sh — Single CLI with extract/distill/run/status/report/review commands
• Learning velocity tracker (JSON) with domain breakdown and streak tracking
• Cron template for automated daily synthesis
• Integration with AGENTS.md, SOUL.md, and daily logs
• Weekly growth reports with plateau detection
• Principle proposal system with human approval gates

What Makes This Different

Every agent can be told "learn from your mistakes." That's a prompt. This is infrastructure:

• Structured extraction for specific signal types
• Typed outputs that go to different files
• Velocity tracking measures rate of learning
• Approval gates for identity-shaping principles
• Plateau detection knows when to go deeper
• 90 days of learning history

The goal isn't perfection. It's compound improvement. 1% better per session, sustained over months.