Global Learning Specification

"Learning answers: What should the agent do differently next time?"

Global Learning provides cross-project, evidence-based wisdom for AI coding agents. Unlike raw scratch notes, lessons must be mathematically substantiated by observed developer experience.

The Learning Lifecycle

1. Candidate: Recorded with initial confidence of 0.5.
2. Evidence: Observations of success or failure are attached.
3. Confidence Recalculation: Confidence is updated via success ratio.
4. Validated / Rejected: Promoted to validated if confidence ≥ 70% with ≥ 3 validations; rejected lessons are pruned.

Confidence Formula

Confidence is calculated deterministically:

confidence = max(0.1, successes / (successes + failures))

CLI Usage

# Record candidate lesson
centr learn add --lesson "Always use Buffer.from(x, 'base64url') when decoding JWTs" --trigger "JWT" --action "Use base64url encoding"

# Add positive evidence
centr learn evidence 1 --outcome success --experience "Unit tests pass"

# List validated learnings
centr learn list --validated