claude-cache 0.2.0

Claude Cache - Memory for your AI coding assistant

Claude Cache

Memory for your AI coding assistant

Claude Cache automatically processes your Claude Code logs, identifies successful patterns, and builds memory that reduces AI hallucinations and improves coding effectiveness.

Note: This is an independent tool for enhancing Claude Code, not an official Anthropic product.

🚨 Important: How Claude Cache Learns

Claude Cache learns from your feedback! When Claude helps you successfully, you must tell it:

• Say "Perfect!", "That worked!", or "Thanks!" when things work
• Without feedback, Cache won't know what succeeded and won't save patterns
• More feedback = Better memory = Smarter Claude

Example:

You: "Add user authentication"
Claude: *implements auth*
You: "Perfect! The login works now" ← This triggers learning!

Features

• First-Run Documentation Import - Scans your entire Development folder for existing documentation
• Automatic Log Processing - Monitors and processes Claude Code session logs in real-time
• Intelligent Lesson Organization - Categorizes lessons by topic (auth, database, API, etc.)
• Smart Document Management - Keeps CLAUDE.md under 30KB with overflow handling
• Success Pattern Detection - Identifies which approaches worked well for specific tasks
• Context Generation - Creates hierarchical context with category-specific files
• Slash Commands - Generates Claude Code slash commands for quick access to patterns
• Convention Tracking - Learns your project's coding conventions and patterns
• Team Knowledge Sharing - Export and import patterns for team collaboration
• User Content Preservation - Your custom notes in CLAUDE.md are never overwritten

How It Works

1. First Run - Prompts to scan your Development folder for all existing documentation
2. Import - Extracts lessons learned, warnings, and best practices from your docs
3. Organize - Categories lessons into topics (authentication, database, API, etc.)
4. Monitor - Watches your Claude Code log files (~/.claude/projects/)
5. Analyze - Detects successful patterns using multiple indicators
6. Store - Builds a searchable knowledge base with SQLite
7. Generate - Creates intelligent index with category files in .claude/lessons/
8. Learn - Improves Claude's responses with hierarchical context

Multi-Project Support

Claude Cache automatically handles multiple projects without any configuration:

• Automatic Detection - Identifies different projects from Claude Code session logs
• Separate Knowledge Bases - Each project gets its own patterns and context
• Project-Specific Learning - React patterns for your frontend won't mix with Python patterns from your API
• Smart Context Injection - Claude reads the right .claude/CLAUDE.md based on which project you're working in

When you switch between projects in Claude Code, Claude Cache automatically switches context too!

Installation

From PyPI (Recommended)

pip install claude-cache

From Source (Development)

git clone https://github.com/ga1ien/claude-cache.git
cd claude-cache
pip install -e .

Quick Start

First Time Setup

cache start

On first run, you'll see:

🎉 Welcome to Claude Cache!

Would you like to scan for existing documentation?
1. Scan all Claude Code projects (from logs)
2. Scan your Development folder     [Default]
3. Scan a custom directory
4. Skip for now

Choose an option [2]: _

This imports all your existing documentation, giving you an immediate knowledge base!

Running Claude Cache

Option 1: Terminal Tab (Simplest)

cache start
# Keep this terminal tab open

Option 2: tmux (Recommended for long sessions)

# Install tmux (one-time)
brew install tmux

# Start in tmux
tmux new -s cache
cache start
# Detach: Ctrl+B then D
# Reattach: tmux attach -t cache

Note: The --daemon flag has issues on macOS. Use tmux or a terminal tab instead.

2. Use Claude Code as normal - Claude Cache will automatically process your interactions

3. Give feedback when things work - Say "Perfect!", "That worked", or "Thanks!" to help Claude Cache learn

4. Access your patterns via slash commands in Claude Code:

• /project-context [task description] - Get relevant patterns for your current task
• /best-practices - View successful approaches
• /conventions - See detected project conventions
• /quick-ref - Quick reference for frequently used files
• /debug-helper - Get debugging assistance

Usage

Basic Commands

# Start monitoring and processing logs
cache start

# Process existing logs without monitoring
cache process

# Query patterns from your knowledge base
cache query "implement authentication"

# Show statistics
cache stats

# Generate slash commands for a project
cache generate --project my-project

Documentation Commands

# Scan repository for documentation (lessons learned, READMEs, etc.)
cache scan-docs /path/to/repo

# Scan current directory
cache scan-docs .

# Search through indexed documentation
cache search-docs --query "authentication" --project my-app

Advanced Usage

# Export patterns for backup or sharing
cache export patterns.json --project my-project

# Import patterns from team member
cache import team-patterns.json

# Rebuild knowledge base from scratch
cache rebuild --confirm

# Generate context for specific request
cache context "fix login bug" --project my-app

Configuration

Claude Cache stores its database at:

~/.claude/knowledge/cache.db

Slash commands are generated in your project's .claude/commands/ directory.

Fine-Tuning Detection

The tool automatically detects and adapts to different tech stacks:

Built-in Stack Detection:

• Frontend: React, Vue, Angular, CSS, Tailwind
• Backend: APIs, Express, Django, Rails
• Database: SQL, migrations, ORMs, query optimization
• Mobile: React Native, Expo, Flutter
• DevOps: Docker, Kubernetes, CI/CD
• Testing: Jest, Pytest, Cypress

Customization via config.yaml:

# Adjust success weights for your stack
stack_weights:
  frontend:
    user_satisfied: 0.3
    no_console_errors: 0.25
    component_renders: 0.25

# Add custom success patterns
custom_patterns:
  react_success:
    keywords:
      - "useEffect working"
      - "state updated correctly"
    weight: 1.3

Per-Project Configuration:

projects:
  my-frontend-app:
    primary_stack: frontend
    success_threshold: 0.75
    extra_patterns:
      - "lighthouse score improved"

The tool learns what "success" means for YOUR specific tech stack and coding patterns.

Adding Custom Tech Stacks

Any tech stack can be added! Edit config.yaml:

custom_stacks:
  your_stack_name:
    keywords: ["your", "stack", "keywords"]
    file_patterns: ["*.ext", "folder/*"]
    success_indicators:
      - "compilation successful"
      - "your success condition"
    failure_indicators:
      - "your error pattern"
    weight_multiplier: 1.3

Examples included for:

• Rust/Actix, Elixir/Phoenix, Go/Gin
• Svelte/SvelteKit, Game Development (Unity/Godot)
• Data Science (Pandas/PyTorch), Blockchain/Web3
• Infrastructure as Code, Embedded Systems

See custom_stacks_example.yaml for complete examples.

How Success Detection Works

Claude Cache uses multiple signals to identify successful patterns:

Your Feedback (Most Important!)

• User satisfaction - "Perfect!", "That worked!", "Thanks!"
• Task completion - "Fixed!", "Done!", "Working now!"

Automatic Detection

• Test Results - "Tests passed", "Build successful" in output
• Error-Free Execution - No exceptions after changes
• File Modifications - Successful edits without errors
• Success Markers - ✓, ✅, "SUCCESS" in output

Sessions scoring above 70% are stored as successful patterns.

Pattern Matching

When you make a new request, the system:

1. Searches for similar previous requests using TF-IDF vectorization
2. Ranks patterns by similarity and success score
3. Generates context from the most relevant patterns
4. Injects this context to improve Claude's response

Project Structure

claude-cache/
├── src/
│   └── cache_for_claude/
│       ├── __init__.py
│       ├── agent.py             # Main coordination agent
│       ├── cli.py               # Command-line interface
│       ├── context_injector.py  # Context and command generation
│       ├── doc_scanner.py       # Documentation scanner for imports
│       ├── lesson_organizer.py  # Categorizes lessons by topic
│       ├── knowledge_base.py    # Database storage
│       ├── log_processor.py     # Log parsing
│       ├── log_watcher.py       # File monitoring
│       └── success_detector.py  # Pattern detection
├── pyproject.toml
└── README.md

Privacy & Security

• All data is stored locally on your machine
• No data is sent to external services
• Sensitive information in logs remains private
• Export/import allows controlled sharing

Use Cases

For Individual Developers

• Build a personal knowledge base of what works
• Reduce repetitive explanations to Claude
• Speed up problem-solving with historical context
• Track your most successful coding patterns

For Teams

• Share successful patterns across team members
• Maintain consistent approaches to common tasks
• Onboard new developers with team-specific knowledge
• Build collective intelligence from everyone's interactions

For Projects

• Document project-specific conventions automatically
• Track which files are frequently modified together
• Identify common debugging patterns
• Maintain project-specific context

Working with Multiple Projects

Claude Cache seamlessly handles multiple projects. Here's what happens:

Automatic Project Separation

# When you work on different projects:
~/Development/my-react-app     → Patterns stored under "my-react-app"
~/Development/python-api        → Patterns stored under "python-api"
~/Development/mobile-app        → Patterns stored under "mobile-app"

Project-Specific Files Created

each-project/
├── .claude/
│   ├── CLAUDE.md              # Main index (5-10KB, Claude reads this)
│   ├── lessons/               # Categorized lessons (unlimited)
│   │   ├── authentication_lessons.md
│   │   ├── database_lessons.md
│   │   ├── api_lessons.md
│   │   └── debugging_lessons.md
│   └── commands/              # Slash commands for Claude Code

Intelligent Document Organization

CLAUDE.md Structure (Kept Small):

• Top 5 critical warnings
• Category index with file paths
• Instructions for Claude
• User notes section (preserved)

Lessons Directory:

• Organized by topic
• 10 lessons per priority level
• Automatic overflow files for >40 lessons
• Smart navigation between parts

Managing Multiple Projects

# View all projects
cache stats

# View specific project
cache stats --project "my-react-app"

# Export one project's patterns
cache export react-patterns.json --project "my-react-app"

# Query across all projects
cache query "authentication"

# Query one project only
cache query "authentication" --project "python-api"

How Claude Uses Project Context

1. You open my-react-app in Cursor/Claude Code
2. Claude automatically reads my-react-app/.claude/CLAUDE.md
3. Claude now knows your React patterns, not your Python patterns
4. Switch to python-api → Claude switches context automatically

No configuration needed - it just works!

Contributing

Contributions are welcome! Please see our Contributing Guide for details.

Roadmap

• Web dashboard for pattern visualization
• Integration with other AI coding tools
• Pattern clustering and categorization
• Automated pattern quality scoring
• Multi-project pattern correlation
• VSCode extension for pattern access

License

MIT License - see LICENSE file for details.

Acknowledgments

Inspired by the insight that "if you document along the way, AI hallucinates less because it has more documentation that's personalized to the project."

Support

• Issues: GitHub Issues
• Discussions: GitHub Discussions
• Documentation: Full Documentation

---

Note: This tool is not officially affiliated with Anthropic or Claude. It's a community tool designed to enhance your Claude Code experience.