Long-Term Memory

About

Long-Term Memory is a drop-in MCP server designed to equip AI assistants, such as Claude and other MCP clients, with advanced persistent memory. It tackles common AI memory challenges like manual retrieval and token waste by automatically injecting relevant memories and intelligently composing top-K memories within specified token budgets. Through a sophisticated combination of semantic search, knowledge graphs, and reinforcement learning, it enables AI agents to retain, retrieve, and organize information effectively across conversations.

Key Features

• RL-powered policy to intelligently decide when to save, skip, or retrieve memories.
• Hybrid retrieval combining semantic search with knowledge graph traversal for comprehensive context.
• Real-time WebSocket-based visualization of the memory graph, showing live updates as memories are processed.
• 0 GitHub stars
• Multi-resolution text composition to optimize token usage by selecting relevant memory summaries or triples.
• Automated memory management including auto-linking, a decay-based forgetting pipeline, and periodic maintenance.

Use Cases

• Automating the retrieval of relevant information for AI agents, reducing the need for explicit user prompts.
• Optimizing LLM context windows by intelligently selecting and composing memories within a token budget.
• Enhancing AI assistants like Claude and OpenClaw with persistent memory across multiple conversations.