muninn

Muninn is an AI memory tool that lets your coding agent retain context across sessions. It records what was decided, what was learned, and what next steps were identified, so your AI can build on prior work rather than starting from zero each time. The MCP tooling exposes a set of memory-related commands (such as muninn_query, muninn_check, muninn_predict, muninn_suggest, muninn_file_add, muninn_decision_add, muninn_learn_add, muninn_issue, muninn_session, muninn_enrich, muninn_approve, muninn_context, muninn_intent, and a general muninn passthrough) that you can integrate with compatible editors and environments. You can register Muninn with your editor using the provided mcp add commands and then run Muninn to begin a session that auto-initializes for your project.

Once set up, Muninn tracks memory types like files, decisions, issues, learnings, sessions, and blast radius to understand what changes influence other parts of your codebase. The 2000-token budget per tool call ensures that only relevant context is surfaced, reducing noise while maintaining useful cross-session insights. You’ll see how changes ripple through the project and how patterns learned in one session inform future edits, leading to smarter, more context-aware AI-assisted development.

Prerequisites:

• Node.js and npm (or use Bun as in the Muninn install guide)
• Basic familiarity with shell commands

Option 1 — npx (Recommended):

npx muninn-ai

Option 2 — Manual Install (requires Bun):

# Clone the repository to a local location
git clone https://github.com/ravnltd/muninn.git ~/.local/share/muninn

# Install and set up the Muninn MCP server wrapper
cd ~/.local/share/muninn && ./install.sh

Register Muninn with your editor (example for Claude Code):

claude mcp add --scope user muninn -- muninn-mcp

Verify installation:

muninn --help
muninn status

If you chose Option 2, ensure ~/.local/bin/ contains wrapper scripts and that they are in your PATH so your MCP client can invoke muninn-mcp as needed.

Tips and common issues:

• Muninn operates with a hub-and-spoke memory model suitable for multi-machine setups. If you’re running in a distributed environment, consider the hub-and-spoke configuration and follow the Multi-Machine Setup guidance in the README.
• The MCP toolset includes a wide range of commands for memory management and project-wide context. Familiarize yourself with muninn_query, muninn_check, muninn_predict, muninn_suggest, and muninn_file_add to maximize memory usefulness.
• If you encounter integration issues with editors, ensure the muninn-mcp script is discoverable in your PATH and that your editor’s MCP integration points to it.
• Muninn’s fragility scoring and budget concepts are designed to minimize context bloat. If you need more or less context, you can adjust which tools you invoke and how you surface context in your workflow.
• For multi-machine setups, export MUNINN_MODE and MUNINN_PRIMARY_URL as described in the documentation to enable HTTP-based coordination across machines.