Kontxt - Local-First MCP Server for AI Memory Management

Remember Everything. Confabulate Nothing.

Kontxt is a local-first MCP (Model Context Protocol) server that solves the AI memory crisis through intelligent context engineering. Built on the proven architecture that achieved 94% task completion rates in production, Kontxt ensures AI agents never forget critical context.

🚀 Quick Start

Installation

npm install -g kontxt-mcp

Configure with Claude Desktop

Add to your Claude Desktop configuration:

Windows: %APPDATA%\Claude\claude_desktop_config.json Mac: ~/Library/Application Support/Claude/claude_desktop_config.json

{
  "mcpServers": {
    "kontxt": {
      "command": "kontxt-mcp"
    }
  }
}

First Use

1. In Claude Desktop, ask the AI to initialize Kontxt:

   Please initialize Kontxt in my project directory

Testing

To test Kontxt without Claude Desktop:

cd kontxt
node test-manual.js

This creates test files in the ../Kontxt Testing/ directory, keeping your project clean.

2. The AI will create the 4-file structure:

   • Implementation.md - Task tracking with checkboxes
   • project_structure.md - Architecture decisions
   • UI_UX_doc.md - Design specifications
   • Bug_tracking.md - Error memory

3. Create your first task:

   Create a task to implement JWT authentication in /api/auth

🧠 Core Concept: Context Engineering

Every task you create is a prompt for a future AI with ZERO context. Kontxt enforces this principle through:

• Quality Assessment: Real-time scoring of task context completeness
• Educational Feedback: Gentle guidance to improve context quality
• Memory Verification: Prevent AI confabulation with ground truth checking
• Context Snapshots: Preserve state at key milestones

📋 Available Tools

init_project

Initialize the Kontxt 4-file structure in your project.

create_task

Create a new task with comprehensive context. The AI will guide you to include:

• File paths and locations
• Error messages or symptoms
• Success criteria
• Technical context

update_task

Mark tasks as pending, in-progress, or completed.

get_current_task

Get the next task to work on with full context.

get_progress

View project progress with context quality metrics.

create_snapshot

Preserve current state for future reference.

verify_memory

Check AI's understanding against ground truth.

pack_context

Optimize context for the current task within token limits.

🎯 Best Practices

1. Write Detailed Tasks: Include file paths, error messages, and success criteria
2. Regular Snapshots: Create snapshots after completing major features
3. Verify Memory: Use memory verification when switching between AI sessions
4. Trust the Guidance: The educational feedback helps you write better tasks

🏗️ Architecture

/your-project
├── .kontxt/                  # Kontxt metadata
│   ├── config.json          # Configuration
│   └── context/             # Snapshots
├── Implementation.md        # Task tracking
├── project_structure.md     # Architecture
├── UI_UX_doc.md            # Design specs
└── Bug_tracking.md         # Error log

🤝 Contributing

Kontxt is open source and welcomes contributions. The codebase is intentionally simple:

• Pure JavaScript (no TypeScript)
• Minimal dependencies
• Clear file structure
• Comprehensive comments

📜 License

MIT License - See LICENSE file for details.

🙏 Acknowledgments

• Andrej Karpathy for articulating "context engineering" as a field
• The AI community for collectively recognizing the importance of persistent context
• DevPartner beta users whose feedback shaped these patterns

Built with the hard-won insights from DevPartner's journey to solve AI memory loss.