Continuous Learning

Continuous Learning v2 - Instinct-Based Architecture

An advanced learning system that turns your Claude Code sessions into reusable knowledge through atomic "instincts" - small learned behaviors with confidence scoring.

What's New in v2

The Instinct Model

An instinct is a small learned behavior:

---
id: prefer-functional-style
trigger: "when writing new functions"
confidence: 0.7
domain: "code-style"
source: "session-observation"
---

# Prefer Functional Style

## Action
Use functional patterns over classes when appropriate.

## Evidence
- Observed 5 instances of functional pattern preference
- User corrected class-based approach to functional on 2025-01-15

Properties:

• Atomic — one trigger, one action
• Confidence-weighted — 0.3 = tentative, 0.9 = near certain
• Domain-tagged — code-style, testing, git, debugging, workflow, etc.
• Evidence-backed — tracks what observations created it

How It Works

Session Activity
      │
      │ Hooks capture prompts + tool use (100% reliable)
      ▼
┌─────────────────────────────────────────┐
│         observations.jsonl              │
│   (prompts, tool calls, outcomes)       │
└─────────────────────────────────────────┘
      │
      │ Observer agent reads (background, Haiku)
      ▼
┌─────────────────────────────────────────┐
│          PATTERN DETECTION              │
│   • User corrections → instinct         │
│   • Error resolutions → instinct        │
│   • Repeated workflows → instinct       │
└─────────────────────────────────────────┘
      │
      │ Creates/updates
      ▼
┌─────────────────────────────────────────┐
│         instincts/personal/             │
│   • prefer-functional.md (0.7)          │
│   • always-test-first.md (0.9)          │
│   • use-zod-validation.md (0.6)         │
└─────────────────────────────────────────┘
      │
      │ /evolve clusters
      ▼
┌─────────────────────────────────────────┐
│              evolved/                   │
│   • commands/new-feature.md             │
│   • skills/testing-workflow.md          │
│   • agents/refactor-specialist.md       │
└─────────────────────────────────────────┘

Quick Start

1. Enable Observation Hooks

Add to your ~/.claude/settings.json.

If installed as a plugin (recommended):

{
  "hooks": {
    "PreToolUse": [{
      "matcher": "*",
      "hooks": [{
        "type": "command",
        "command": "${CLAUDE_PLUGIN_ROOT}/skills/continuous-learning/hooks/observe.sh pre"
      }]
    }],
    "PostToolUse": [{
      "matcher": "*",
      "hooks": [{
        "type": "command",
        "command": "${CLAUDE_PLUGIN_ROOT}/skills/continuous-learning/hooks/observe.sh post"
      }]
    }]
  }
}

If installed manually to ~/.claude/skills:

{
  "hooks": {
    "PreToolUse": [{
      "matcher": "*",
      "hooks": [{
        "type": "command",
        "command": "~/.claude/skills/continuous-learning/hooks/observe.sh pre"
      }]
    }],
    "PostToolUse": [{
      "matcher": "*",
      "hooks": [{
        "type": "command",
        "command": "~/.claude/skills/continuous-learning/hooks/observe.sh post"
      }]
    }]
  }
}

2. Initialize Directory Structure

The Python CLI will create these automatically, but you can also create them manually:

mkdir -p ~/.claude/homunculus/{instincts/{personal,inherited},evolved/{agents,skills,commands}}
touch ~/.claude/homunculus/observations.jsonl

3. Use the Instinct Commands

/instinct-status     # Show learned instincts with confidence scores
/evolve              # Cluster related instincts into skills/commands
/instinct-export     # Export instincts for sharing
/instinct-import     # Import instincts from others

Commands

Configuration

Edit config.json:

{
  "version": "2.0",
  "observation": {
    "enabled": true,
    "store_path": "~/.claude/homunculus/observations.jsonl",
    "max_file_size_mb": 10,
    "archive_after_days": 7
  },
  "instincts": {
    "personal_path": "~/.claude/homunculus/instincts/personal/",
    "inherited_path": "~/.claude/homunculus/instincts/inherited/",
    "min_confidence": 0.3,
    "auto_approve_threshold": 0.7,
    "confidence_decay_rate": 0.05
  },
  "clarifications": {
    "enabled": true,
    "clarifications_path": "~/.claude/homunculus/clarifications/",
    "min_confidence": 0.5,
    "detection_patterns": [
      "rapid_re_edit",
      "error_then_success",
      "repeated_tool",
      "iterative_correction"
    ],
    "analysis_model": "haiku",
    "max_clarifications": 50
  },
  "observer": {
    "enabled": true,
    "model": "haiku",
    "run_interval_minutes": 5,
    "patterns_to_detect": [
      "user_corrections",
      "error_resolutions",
      "repeated_workflows",
      "tool_preferences"
    ]
  },
  "evolution": {
    "cluster_threshold": 3,
    "evolved_path": "~/.claude/homunculus/evolved/"
  }
}

Real-time Clarification Detection (NEW)

v2 includes automatic misunderstanding detection that captures communication issues as they happen:

How It Works

When the hook system detects potential corrections (rapid re-edits, error-recovery patterns, repeated tool usage), it triggers a lightweight background analysis:

Correction Pattern Detected
         ↓
Background Analysis (Haiku)
  - Analyzes last 5 turns only (~1-2K tokens)
  - Generates clarification hint (50 words)
         ↓
Saves to clarifications/

Detection Patterns

• Rapid Re-edit: Same file edited multiple times quickly
• Error Recovery: Error followed by successful fix
• Repeated Tool: Same tool used 3+ times in sequence
• Iterative Correction: Quick back-and-forth on a task

Example Clarification

---
id: refactor-scope-20250204
trigger: "when editing files multiple times"
confidence: 0.70
type: disambiguation
---

When user says "refactor", clarify: "Structural change or rename only?"

Evidence:
- Rapid tool iteration: Edit → Edit → Edit
- Detected by hook: rapid_re_edit

Token Efficiency

Traditional approach: Analyze entire session (10-20K tokens) New approach: Analyze last 5 turns only (1-2K tokens) Efficiency gain: 5-10x reduction

File Structure

~/.claude/homunculus/
├── identity.json           # Your profile, technical level
├── observations.jsonl      # Current session observations
├── observations.archive/   # Processed observations
├── clarifications/         # Real-time disambiguation hints (NEW)
├── instincts/
│   ├── personal/           # Auto-learned instincts
│   └── inherited/          # Imported from others
└── evolved/
    ├── agents/             # Generated specialist agents
    ├── skills/             # Generated skills
    └── commands/           # Generated commands

Integration with Skill Creator

When you use the Skill Creator GitHub App, it now generates both:

• Traditional SKILL.md files (for backward compatibility)
• Instinct collections (for v2 learning system)

Instincts from repo analysis have source: "repo-analysis" and include the source repository URL.

Confidence Scoring

Confidence evolves over time:

Confidence increases when:

• Pattern is repeatedly observed
• User doesn't correct the suggested behavior
• Similar instincts from other sources agree

Confidence decreases when:

• User explicitly corrects the behavior
• Pattern isn't observed for extended periods
• Contradicting evidence appears

Why Hooks vs Skills for Observation?

"v1 relied on skills to observe. Skills are probabilistic—they fire ~50-80% of the time based on Claude's judgment."

Hooks fire 100% of the time, deterministically. This means:

• Every tool call is observed
• No patterns are missed
• Learning is comprehensive

Backward Compatibility

v2 is fully compatible with v1:

• Existing ~/.claude/skills/learned/ skills still work
• Stop hook still runs (but now also feeds into v2)
• Gradual migration path: run both in parallel

Privacy

• Observations stay local on your machine
• Only instincts (patterns) can be exported
• No actual code or conversation content is shared
• You control what gets exported

Related

• Skill Creator - Generate instincts from repo history
• Homunculus - Inspiration for v2 architecture
• The Longform Guide - Continuous learning section

Instinct-based learning: teaching Claude your patterns, one observation at a time.