Forgetful is an open-source Model Context Protocol (MCP) server designed to provide persistent storage and retrieval for AI agents. It addresses the critical need for shared knowledge bases, enabling agents to access past information and maintain context across interactions, thereby enhancing their effectiveness, particularly in complex workflows like software development. It targets developers and researchers building or integrating AI agents that require robust memory capabilities.

How It Works

Forgetful adopts an opinionated Zettelkasten-inspired approach, enforcing atomic memories (one concept per note) with title, content, context, keywords, and tags. It generates semantic embeddings for natural language retrieval and automatically links semantically similar memories, facilitating the construction of a knowledge graph. The system also manages structured data like entities (people, organizations), projects, documents, code artifacts, skills, and plans for multi-agent coordination, aiming to improve recall accuracy.

Quick Start & Requirements

• Installation: Recommended via PyPI (uvx forgetful-ai or uv tool install forgetful-ai). Source installation and Docker deployments (SQLite or PostgreSQL) are also supported.
• Transport: Defaults to stdio; HTTP transport is available (--transport http --port 8020).
• Dependencies: Requires uvx or uv for execution. Docker requires Docker Compose.
• Data Storage: Uses platform-appropriate locations by default (e.g., ~/.local/share/forgetful).
• Links: Configuration Guide, Connectivity Guide, Complete Tool Reference.

Highlighted Details

• Supports STDIO and HTTP transport mechanisms.
• Offers flexible storage: SQLite (default) and PostgreSQL (for scale).
• Enables vector-based retrieval via natural language queries, enhanced by cross-encoder reranking.
• Features automatic linking of semantically similar memories for knowledge graph construction.
• Provides robust support for multi-agent coordination (Plans/Tasks) and procedural knowledge (Skills with SKILL.md import/export).
• Facilitates local embedding generation via FastEmbed, avoiding cloud dependencies.

Maintenance & Community

The project welcomes contributions, offering detailed guides for testing, development setup, and CI/CD. A roadmap outlines planned features. Specific community channels (e.g., Discord, Slack) are not explicitly mentioned.

Licensing & Compatibility

Released under the MIT License, permitting commercial use and closed-source linking.

Limitations & Caveats

The mandatory atomic memory structure may require agents to decompose complex information. Token budget management prioritizes results by importance and recency, potentially truncating less critical context to prevent LLM context window overflow.