/compact

Deduplicate and merge learnings files using semantic analysis.

Process

1. Read All Learnings

cat ~/.claude/learnings/failures.jsonl 2>/dev/null
cat ~/.claude/learnings/discoveries.jsonl 2>/dev/null
cat ~/.claude/learnings/constraints.jsonl 2>/dev/null

If files are empty or missing, report "No learnings to compact" and exit.

2. Semantic Analysis (LLM)

For each category, analyze entries and identify:

Duplicates to merge:

• Same insight expressed differently
• Same root cause for different symptoms
• Overlapping applies_to tags with same discovery

Merge strategy:

• Keep most specific/detailed wording
• Boost confidence: +0.05 per merged entry (max 0.95)
• Combine applies_to arrays (union)
• Use most recent timestamp

Entries to prune:

• Outdated (superseded by newer, more accurate learning)
• Too specific to a one-time situation
• Temporary constraints that no longer apply
• Low confidence (< 0.5) with no reinforcement

DO NOT use fixed rules like "90 days old = prune". Assess each entry's current relevance semantically.

3. Create Backup

Before writing changes:

BACKUP_DIR=~/.claude/learnings/backup-$(date +%Y%m%d-%H%M%S)
mkdir -p "$BACKUP_DIR"
cp ~/.claude/learnings/*.jsonl "$BACKUP_DIR/" 2>/dev/null

4. Write Compacted Files

For each category, write the compacted JSONL:

# Clear and rewrite (atomic operation)
cat > ~/.claude/learnings/discoveries.jsonl << 'JSONL'
{"ts":"...","context":"...","discovery":"...","confidence":0.X,"applies_to":["..."]}
...
JSONL

5. Report Changes with Rationale

Output a detailed report explaining decisions:

## Compaction Report

Backup created: ~/.claude/learnings/backup-YYYYMMDD-HHMMSS/

### discoveries.jsonl: N → M entries

**Merged:**
- "insight A" + "insight B" → "merged insight" (confidence 0.80 → 0.85)
  - Reason: Same discovery about [topic], different wording

**Pruned:**
- "old specific insight"
  - Reason: Superseded by more general learning

**Kept:** X entries unchanged

### failures.jsonl: N → M entries

[Similar format]

### constraints.jsonl: N → M entries

[Similar format]

---
Total: X entries → Y entries (Z merged, W pruned)

Rules

• LLM judges all decisions - No hardcoded thresholds
• Always backup first - Never lose data
• Explain every change - Transparency in reasoning
• Conservative pruning - When in doubt, keep
• Quality over quantity - Fewer precise learnings > many vague ones

Limits

• Max 50 entries per category after compaction
• If over limit, prioritize by: confidence > recency > specificity