Preferences inferred from user corrections and behavior patterns. Structured YAML with categories matching stated preferences. WRITE when user corrects agent behavior (e.g., edits a draft, rejects a suggestion). Each entry includes: observation, inferred preference, confidence (low/medium/high), and supporting evidence count. Lower confidence than stated—these are inferences. If user corrects the same thing 3+ times, consider it high confidence. PROMOTE to stated preferences if user confirms. DELETE if contradicted.