# Memory System Guide

SuperClaude Framework includes a built-in memory system called **ReflexionMemory** that helps the PM Agent learn from past mistakes and avoid repeating errors.

## Overview

ReflexionMemory is an automatic error learning system that:
- **Remembers** past errors and their solutions
- **Learns** from mistakes to prevent recurrence
- **Searches** for similar errors when new problems occur
- **Persists** across sessions via local file storage

**Key Point**: ReflexionMemory is built-in and requires **no installation or setup**. It works automatically.

## How It Works

### 1. Automatic Error Detection

When the PM Agent encounters an error during implementation:

```yaml
Error Occurs:
  → PM Agent detects failure
  → Analyzes root cause
  → Documents the learning
  → Saves to ReflexionMemory
```

### 2. Learning Storage

Each error is stored as a "reflexion entry" containing:

| Field | Description | Example |
|-------|-------------|---------|
| `task` | What you were trying to do | `"implement JWT authentication"` |
| `mistake` | What went wrong | `"JWT validation failed"` |
| `evidence` | Proof of the error | `"TypeError: Cannot read property 'verify'"` |
| `rule` | Lesson learned | `"Always check environment variables before implementation"` |
| `fix` | How it was solved | `"Added SUPABASE_JWT_SECRET to .env"` |
| `tests` | Verification steps | `["Check .env.example", "Verify all env vars set"]` |
| `status` | Current state | `"adopted"` (active rule) |

### 3. Smart Error Lookup

Next time a similar error occurs:

```yaml
New Error:
  1. ReflexionMemory searches past errors
  2. Finds similar mistakes (keyword matching)
  3. Returns known solutions
  4. PM Agent applies fix immediately

Result:
  ✅ Instant resolution
  ✅ Zero additional tokens
  ✅ <10% error recurrence rate
```

## Storage Location

ReflexionMemory data is stored locally in your project:

```
<project-root>/
└── docs/
    └── memory/
        └── reflexion.jsonl    # Error learning database
```

**Format**: [JSON Lines](https://jsonlines.org/) - one JSON object per line

**Persistence**: Persists across sessions, commits, and branches

## Search Algorithm

ReflexionMemory uses **keyword-based similarity matching**:

1. Extract keywords from current error message
2. Compare with keywords from past errors
3. Calculate overlap ratio: `overlap = (matching_keywords) / (total_keywords)`
4. Return entries with >50% overlap
5. Sort by timestamp (most recent first)

**Example**:
```python
Current error: "JWT token validation failed missing secret"
Past error:    "JWT validation failed secret not found"
Overlap:       7/8 keywords match = 87.5% similarity ✅
```

## User Interaction

### Fully Automatic (Default)

ReflexionMemory works transparently:
- ✅ Auto-loads at session start
- ✅ Auto-searches when errors occur
- ✅ Auto-saves new learnings
- ✅ No explicit commands needed

### Manual Inspection (Optional)

You can view the memory file directly:

```bash
# View all learnings
cat docs/memory/reflexion.jsonl | jq

# Search for specific topic
cat docs/memory/reflexion.jsonl | jq 'select(.task | contains("auth"))'

# Count total learnings
wc -l docs/memory/reflexion.jsonl
```

### Managing Entries

**Clear all memory** (use with caution):
```bash
rm docs/memory/reflexion.jsonl
```

**Remove specific entry**: Edit the file manually and delete the line

**Mark as obsolete**: Change `"status": "adopted"` to `"status": "deprecated"`

## Integration with PM Agent

### When It's Used

ReflexionMemory activates during:

1. **Error Recovery**: When implementation fails
2. **Pre-Implementation**: Checking for known pitfalls
3. **Root Cause Analysis**: Investigating systemic issues

### Workflow Example

```yaml
Scenario: User asks to implement OAuth login

Step 1 - Pre-Check:
  PM Agent: "Checking past OAuth implementations..."
  ReflexionMemory: Found 2 similar tasks
  PM Agent: "⚠️ Warning: Past mistake - forgot to set OAUTH_SECRET"

Step 2 - Implementation:
  PM Agent: Implements OAuth + remembers to check env vars
  Result: Success on first try ✅

Step 3 - If Error Occurs:
  PM Agent: "Error: OAUTH_REDIRECT_URL not configured"
  ReflexionMemory: No similar error found
  PM Agent: Investigates, fixes, documents learning
  ReflexionMemory: Saves new entry for future reference
```

## Performance Benefits

### Token Efficiency

- **With ReflexionMemory**: 500 tokens (direct solution lookup)
- **Without Memory**: 2-10K tokens (full investigation needed)
- **Savings**: 75-95% token reduction on known errors

### Time Savings

- **Known errors**: <30 seconds (instant solution)
- **First occurrence**: 5-15 minutes (investigation + learning)
- **Recurrence rate**: <10% (learns from mistakes)

## File Format Reference

See `docs/memory/reflexion.jsonl.example` for sample entries.

Each line is a complete JSON object:

```json
{"ts": "2025-10-30T14:23:45+09:00", "task": "implement auth", "mistake": "JWT validation failed", "evidence": "TypeError: secret undefined", "rule": "Check env vars before auth implementation", "fix": "Added JWT_SECRET to .env", "tests": ["Verify .env vars", "Test JWT signing"], "status": "adopted"}
```

## Future Enhancements

Current: **Keyword-based search** (50% overlap threshold)

Planned: **Semantic search** upgrade
- Use embeddings for similarity
- Support natural language queries
- Achieve 90% token reduction (industry benchmark)
- Optional vector database integration

## Comparison with Other Systems

| Feature | ReflexionMemory | Mindbase (Planned) | Mem0/Letta |
|---------|-----------------|-------------------|------------|
| **Setup** | Built-in | Never implemented | External install |
| **Storage** | Local JSONL | N/A | PostgreSQL/Vector DB |
| **Search** | Keyword (50%) | N/A | Semantic |
| **Scope** | Errors only | N/A | Full memory |
| **Cost** | Free | N/A | Infrastructure |

**Why ReflexionMemory**: Focused, efficient, and requires zero setup.

## Troubleshooting

### Memory file not found

If `docs/memory/reflexion.jsonl` doesn't exist:
- ✅ Normal on first run
- ✅ Created automatically on first error
- ✅ No action needed

### Entries not being used

Check:
1. Is the error really similar? (View entries manually)
2. Is `status: "adopted"`? (Deprecated entries are ignored)
3. Is keyword overlap >50%? (May need more specific error messages)

### File growing too large

ReflexionMemory files rarely exceed 1MB. If needed:
1. Archive old entries: `mv reflexion.jsonl reflexion-archive-2025.jsonl`
2. Keep recent entries: `tail -100 reflexion-archive-2025.jsonl > reflexion.jsonl`

### Corrupted JSON

If you manually edit and break the JSON format:
```bash
# Validate each line
cat docs/memory/reflexion.jsonl | while read line; do echo "$line" | jq . || echo "Invalid: $line"; done

# Remove invalid lines
cat docs/memory/reflexion.jsonl | while read line; do echo "$line" | jq . >/dev/null 2>&1 && echo "$line"; done > fixed.jsonl
mv fixed.jsonl docs/memory/reflexion.jsonl
```

## Best Practices

### For Users

1. **Let it work automatically** - Don't overthink it
2. **Review learnings periodically** - Understand what patterns emerge
3. **Keep error messages clear** - Better search matching
4. **Don't delete blindly** - Old learnings can be valuable

### For Contributors

1. **Use structured error messages** - Consistent keywords improve matching
2. **Document root causes** - Not just symptoms
3. **Include verification steps** - Make fixes reproducible
4. **Mark outdated entries** - Set status to "deprecated" instead of deleting

## Related Documentation

- **Implementation**: `superclaude/core/pm_init/reflexion_memory.py`
- **Research**: `docs/research/reflexion-integration-2025.md`
- **PM Agent Integration**: `superclaude/agents/pm-agent.md`
- **Architecture**: `docs/reference/pm-agent-autonomous-reflection.md`

## Quick Reference

```bash
# View all learnings
cat docs/memory/reflexion.jsonl | jq

# Search for auth-related errors
grep -i "auth" docs/memory/reflexion.jsonl | jq

# Count learnings
wc -l docs/memory/reflexion.jsonl

# Latest 5 errors
tail -5 docs/memory/reflexion.jsonl | jq

# Check for duplicates (same mistake)
cat docs/memory/reflexion.jsonl | jq -r '.mistake' | sort | uniq -c | sort -rn
```

---

**Questions?** See the [FAQ](../FAQ.md) or open an issue on GitHub.