mcp-memory-service

Why we included this project

Multi-step agent pipelines share a weakness: after a run, the next one starts from nothing because the context window only keeps what is currently in front of the model. mcp-memory-service gives those agents a persistent store with a REST API that saves decisions and retrieves them in about 5ms, and it works with LangGraph, CrewAI, AutoGen, or plain Claude without requiring an MCP client library. The knowledge graph preserves causal relationships between facts, which helps with debugging and lets several agents coordinate around shared state. Everything runs on your own infrastructure, including the ONNX embeddings, so you are not paying per token to a third-party memory API. If you want durable, shared memory across agents and Claude sessions without cloud lock-in, this is a straightforward fit.