recall

# Recall Learnings

Search ~/.claude/learnings/ for entries matching: $ARGUMENTS

## When to Recall

- **Before starting substantial work** → Recall domain-specific learnings
- **When entering familiar territory** → "What did I learn last time?"
- **Before delegation** → Check delegation.jsonl for past patterns
- **After a failure** → Check if this mistake was already captured

## Task

1. Read these files (skip silently if missing):
   - `~/.claude/learnings/failures.jsonl`
   - `~/.claude/learnings/discoveries.jsonl`
   - `~/.claude/learnings/constraints.jsonl`
   - `~/.claude/learnings/delegation.jsonl`

2. Filter entries where `context` or `applies_to` matches the provided hint (or infer from $ARGUMENTS if empty)

3. Prioritize:
   - Recent entries (last 30 days)
   - High-confidence discoveries (>= 0.8)
   - Recurring failures (same pattern 2+ times)

4. Return top 5 most relevant learnings in this format:

### Relevant Failures
- **[context]**: [failure] → Prevention: [prevention]

### Relevant Discoveries
- **[context]** (confidence: X): [discovery]

### Active Constraints
- **[context]**: [constraint] (permanent: yes/no)

### Delegation Learnings
- **[shape_chosen] → [outcome]**: [prevention]

5. If no matches found: "No learnings found for [hint]"

---

## Integration with Loop

When loop:start invokes recall:

1. Extract domain hints from spec (e.g., "auth", "validation", "API")
2. Search all learning categories
3. Surface relevant past failures before decomposition
4. Flag high-confidence constraints that apply

**Automatic invocation:** Loop should call `/hope:recall {domain}` before task decomposition.

---

## Capture Guidance

When you encounter patterns worth capturing, recommend `/hope:learn`:

### What to Capture

| Category | Capture When |
|----------|--------------|
| **Failure** | Same mistake twice, unexpected behavior, debugging took >30 min |
| **Discovery** | Non-obvious solution, library quirk, undocumented behavior |
| **Constraint** | Ha