# Learning Model

## Inputs that count
- explicit corrections
- explicit durable preferences
- self-reflection after meaningful work
- repeated successful workflows

## Inputs that do not count
- silence
- one-off instructions
- guesses about intent
- ambiguous praise with no reusable lesson

## Reflection format
```text
CONTEXT: [task type]
REFLECTION: [what happened]
LESSON: [what to do differently]
```

## Promotion rules
- repeated 3x in 7 days → promote to HOT memory
- unused 30 days → demote to warm storage
- unused 90 days → archive

## Conflict resolution
1. project > domain > global
2. newer > older at same scope
3. if still ambiguous, ask