MARM MCP addresses AI's inherent lack of persistent memory by providing a unified, memory-powered collaborator. It acts as a universal MCP server, enabling AI agents to learn, remember, and share context across sessions and tools, enhancing productivity for developers and researchers.

How It Works

The system utilizes the MARM protocol with a FastAPI backend and SQLite for persistent storage, integrating Sentence Transformers for semantic search and auto-classification. This architecture allows multiple AI agents to share knowledge seamlessly, ensuring long-term recall and session continuity beyond stateless LLMs. It supports HTTP, STDIO, and experimental WebSocket transports.

Quick Start & Requirements

• Primary Install: Docker (docker pull lyellr88/marm-mcp-server:latest then docker run ...) or pip (pip install marm-mcp-server).
• Prerequisites: Python 3.10+.
• Dependencies: FastAPI, Sentence Transformers, SQLite.
• Links: Demo video: https://github.com/user-attachments/assets/c7c6a162-5408-4eda-a461-610b7e713dfe. Installation guides available via links in README (e.g., INSTALL-DOCKER.md, INSTALL-WINDOWS.md).

Highlighted Details

• Facilitates cross-platform AI memory and multi-agent coordination, enabling shared learning among different LLMs.
• Includes 18 MCP tools for comprehensive memory management, session logging, and reasoning.
• Docker-ready with health checks for robust containerized deployment.
• Proven in production for industrial automation and DevOps, demonstrating reliable recall and session management in complex workflows.

Maintenance & Community

The project is actively maintained with planned Q1 2026 improvements in semantic search and session handling. Community engagement is fostered through MARM Discord and GitHub Discussions.

Licensing & Compatibility

• License: MIT License.
• Compatibility: Permits broad use, modification, and distribution, including commercial applications. Derivative works are allowed but must be clearly marked as unofficial.

Limitations & Caveats

WebSocket transport is in beta. Local development authentication is mock-based and not production-ready; multi-user OAuth is a future enhancement. As a project in its early roadmap phase, users should expect ongoing development and potential changes.