learning-engine

name: learning-engine description: "Self-learning system based on SICA, VIGIL, and Mem0 patterns. Auto-triggers after task completion. Captures patterns, consolidates memory, generates skills, recovers from errors." triggers: ["*"] priority: 1 metadata: version: "4.0" research_basis: - "SICA: Self-Improving Coding Agent (arXiv:2504.15228)" - "VIGIL: Reflective Runtime for Self-Healing (arXiv:2512.07094)" - "Mem0: Memory Consolidation Architecture" - "Synapse: Spreading Activation Memory (arXiv:2601.02744)" - "ReCreate: Experience-Driven Agent Creation (arXiv:2601.11100)"

Learning Engine v4.0 (Self-Learning Core)

Overview

The learning engine implements a continuous self-improvement loop:

OBSERVE → REFLECT → PROPOSE → TEST → INTEGRATE
   ↓         ↓         ↓        ↓        ↓
Capture   Analyze   Generate   Rank    Accept if
metrics   failures  options    by      beneficial
& traces  & gaps             merit

Self-Learning Loop (SICA Pattern)

1. Observe

• Capture all tool executions via learning-capture.py hook
• Record decisions, outcomes, durations
• Store episodic traces in ~/.claude/learning/episodes/

2. Reflect

• Analyze failures with VIGIL pattern (error-recovery.py hook)
• Classify errors and suggest recovery strategies
• Generate reflection on what went wrong

3. Propose

• When patterns repeat 5+ times, propose skill generation
• Run scripts/generate-skill.py to create new skills
• Store in ~/.claude/skills/generated/

4. Test

• Validate new patterns against success criteria
• Compare against baseline performance
• Check for regressions

5. Integrate

• Accept improvements if performance > baseline
• Consolidate memory with scripts/consolidate-memory.py
• Update agent capabilities and routing

Quick Classification

After task completion, classify:

1. Task Type: feature | bug | refactor | test | docs | research | security
2. Domain: frontend | backend | database | devops | security | ai-ml | orchestration
3. Complexity: simple | moderate | complex | critical
4. Agent Used: Record which agent handled the task
5. Outcome: success | partial | failed | escalated
6. Learning Value: low | medium | high (errors are high value)

Memory Architecture (Three-Tier)

Tier 1: Working Memory

• Current conversation context
• TTL: 15 minutes
• Storage: In-context
• Purpose: Immediate recall

Tier 2: Episodic Memory

• Full interaction traces with timestamps
• TTL: 30 days
• Storage: ~/.claude/learning/episodes/
• Purpose: Experience replay, pattern extraction

Tier 3: Semantic Memory

• Consolidated facts and patterns
• TTL: Permanent
• Storage: ~/.claude/learning/semantic/
• Purpose: Long-term knowledge

Consolidation Process (Mem0 Pattern)

Run periodically (after 5 interactions or end of session):

python3 ~/.claude/skills/learning-engine/scripts/consolidate-memory.py

Operations:

• ADD: Genuinely new information
• UPDATE: Augment existing with recent details
• DELETE: Remove contradicted facts
• NOOP: Already exists or irrelevant

Dynamic Skill Generation (ReCreate Pattern)

When patterns repeat 5+ times:

python3 ~/.claude/skills/learning-engine/scripts/generate-skill.py

Generated skills stored in ~/.claude/skills/generated/ with:

• SKILL.md definition
• Metadata (confidence, domain, pattern type)
• Auto-incremented instance count

Error Recovery (VIGIL Pattern)

Error taxonomy with recovery strategies:

Gap Detection Signals

• No good agent match found for task
• User corrections needed after completion
• Repeated explanations required
• Task outside known domains
• Low success rate (<70%) on task type
• Multiple escalations needed (3+)
• Confidence below threshold (0.7)

When a gap is detected:

1. Log to ~/.claude/learning/gaps/
2. Suggest agent creation
3. Track for pattern emergence

Storage Structure

~/.claude/learning/
├── episodes/           # Episodic memory (by date)
│   └── YYYY-MM-DD-episodes.jsonl
├── semantic/           # Semantic memory (consolidated)
│   ├── facts.jsonl
│   ├── patterns.jsonl
│   └── decisions.jsonl
├── patterns/           # Extracted patterns (by domain)
│   ├── go/
│   ├── typescript/
│   └── general/
├── solutions/          # Problem-solution pairs
│   ├── error_resolution/
│   └── feature_implementation/
├── gaps/               # Detected capability gaps
│   └── index.json
├── metrics/            # Usage statistics
│   └── index.json
├── errors/             # Error log for VIGIL
│   └── error-log.jsonl
└── consolidated/       # Consolidation history
    └── consolidation-log.jsonl

Pattern Storage Format

{
  "id": "auth-jwt-refresh",
  "domain": "backend",
  "type": "code_pattern",
  "title": "JWT Token Refresh",
  "description": "Pattern for refreshing expired JWT tokens",
  "code_snippet": "...",
  "tags": ["auth", "jwt", "security"],
  "usage_count": 5,
  "success_rate": 0.9,
  "confidence": 0.85,
  "created": "2026-01-15",
  "last_used": "2026-01-26",
  "hash": "abc123def456"
}

Hooks Integration

learning-capture.py (PostToolUse)

• Captures all tool executions
• Extracts patterns from successful interactions
• Updates metrics

error-recovery.py (PostToolUse)

• Analyzes errors with VIGIL pattern
• Suggests recovery strategies
• Logs for future learning

telemetry-collector.py (PostToolUse)

• Tracks routing decisions
• Measures latency and cost
• Feeds into optimization

context-optimizer.py (PreToolUse)

• Applies progressive disclosure
• Manages token budget
• Tracks context usage

Commands

# Run memory consolidation
python3 ~/.claude/skills/learning-engine/scripts/consolidate-memory.py

# Generate skills from patterns
python3 ~/.claude/skills/learning-engine/scripts/generate-skill.py

# View learning metrics
cat ~/.claude/learning/metrics/index.json | jq .

# View recent episodes
tail -20 ~/.claude/learning/episodes/$(date +%Y-%m-%d)-episodes.jsonl | jq .

# View error log
tail -20 ~/.claude/learning/errors/error-log.jsonl | jq .

Performance Targets

Advanced Resources

• @scripts/generate-skill.py - Dynamic skill generation
• @scripts/consolidate-memory.py - Memory consolidation
• @resources/advanced-metrics.md - Detailed metrics guide
• @resources/error-taxonomy.md - Full error classification