Continuous Learning

Pattern Extraction Framework

After every significant coding session, extract and categorize learnings into three buckets:

1. Corrections - Mistakes caught during review or by the user
2. Successful Approaches - Patterns that worked well and should be repeated
3. Anti-Patterns - Approaches that caused problems and should be avoided

Learning Entry Format

pattern:
  id: "LEARN-2025-0042"
  category: "error-handling"
  type: "correction"         # correction | success | anti-pattern
  confidence: 0.85           # 0.0 to 1.0
  language: "typescript"
  context: "API error responses"
  observation: "Returning raw error messages from database exceptions exposes internals"
  lesson: "Always map database errors to application-level error codes before returning"
  example:
    before: "catch (e) { res.status(500).json({ error: e.message }) }"
    after: "catch (e) { logger.error(e); res.status(500).json({ error: 'INTERNAL_ERROR' }) }"
  frequency: 3               # times this pattern has been observed
  last_seen: "2025-06-15"

Confidence Scoring

Score	Meaning	Action
0.95+	Verified across multiple projects	Apply automatically
0.80-0.94	Confirmed in this codebase	Apply and mention
0.60-0.79	Observed but not fully validated	Suggest with caveat
0.40-0.59	Hypothesis based on limited data	Ask before applying
<0.40	Speculative, needs validation	Document but do not apply

Update confidence based on:

• +0.10 when pattern is confirmed correct by user
• +0.05 when pattern is observed again in a different context
• -0.15 when pattern leads to a correction
• -0.20 when pattern is explicitly rejected by user

Session Wrap-Up Protocol

At the end of each session or before context compaction: