Neural Session Train

• Task description
• Approach taken
• Outcome (success, partial, failure)
• Agents involved
• Whether a memory pattern was reused

2. Score Reused Patterns

For each task that used an existing pattern from memory:

1. Identify the pattern ID that was matched
2. Determine the outcome (success/partial/failure)
3. Invoke neural-pattern-score logic to update confidence and usage stats
4. Record the scoring event

3. Extract New Patterns

For each task that did NOT use an existing pattern:

1. Evaluate whether the approach is generalizable (skip one-off fixes)
2. If generalizable, invoke neural-pattern-extract logic to create a new pattern
3. Assign initial confidence of 0.5
4. Add to the index

4. Check for Consolidation Opportunities

After adding new patterns, scan for near-duplicates:

• New pattern tags overlap 80%+ with an existing pattern
• Approach descriptions are semantically similar
• Flag these for consolidation but do not auto-merge during training — let neural-pattern-prune handle it

5. Update Learning Rate

If 5 or more pattern-use events occurred this session:

1. Calculate session accuracy from scoring results
2. Invoke neural-learn-rate logic to adjust the rate if warranted
3. If fewer than 5 events, skip rate adjustment

6. Produce Training Report

## Session Training Report — [timestamp]

### Tasks Reviewed: [N]

### Patterns Reused
| Pattern | Task | Outcome | Confidence Change |
|---------|------|---------|-------------------|
| pat-retry-backoff | Fix timeout errors | success | 0.65 → 0.70 |

### New Patterns Extracted
| Pattern | Source Task | Tags | Confidence |
|---------|-----------|------|-----------|
| pat-docker-layer-cache | Optimize build time | docker, performance | 0.50 |

### Patterns Flagged
- Near-duplicate: pat-new-approach ≈ pat-existing-approach (80% tag overlap)

### Learning Rate
- Session accuracy: [x.xx]
- Rate adjustment: [old] → [new] ([converge/explore/hold])

### Memory Store Summary
- Total patterns: [N]
- Average confidence: [x.xx]
- Patterns above 0.7: [N] (reliable)
- Patterns below 0.2: [N] (candidates for pruning)

7. Save and Log

• Write the training report to state/STATE.md under the session section
• Append a summary to logs/claude_log.md
• Ensure all pattern files and index.json are saved

Output

• Updated pattern files in memory/patterns/ with new scores and entries
• Updated memory/patterns/index.json
• Updated memory/sona-config.json (if learning rate was adjusted)
• Session training report in state/STATE.md and logs/claude_log.md

Review Checklist

• Every completed task in the session was reviewed
• Reused patterns had their confidence and usage updated
• New patterns were only extracted for generalizable approaches
• Near-duplicates were flagged but not auto-merged
• Learning rate was adjusted only if 5+ scoring events occurred
• Training report includes all sections with accurate data
• All memory files are valid JSON after updates