Surreal-Memory: persistent memory for AI agents
Surreal-Memory gives AI agents persistent context across sessions through a Model Context Protocol integration. It is an open-source implementation for teams that need memory to be inspectable, testable and separate from the model itself.
The problem it addresses
An agent can lose relevant decisions, preferences and task history when a session ends. Simply storing every message does not solve that problem: the system still has to decide what is worth retaining and what should be retrieved for the current task. Surreal-Memory provides an explicit memory workflow instead of hiding persistence inside an opaque prompt.
My role and architecture
I am the author, architect and maintainer. I design the memory workflow and develop the implementation in Python, with SurrealDB for persistence and MCP as the integration boundary. The public repository makes the storage model, interfaces and implementation choices available for inspection.
Where it is useful
The project is relevant to MCP integration, long-running agents, context engineering and retrieval diagnostics. It can support workflows that need decisions or preferences to survive between sessions while keeping memory access behind an explicit tool interface.
Verification and limits
Persistence is not the same as truth. A stored fact can become stale, and a relevant memory can still be missed or retrieved in the wrong context. Source checking, scoped recall and independent verification remain part of the workflow. Public source allows those assumptions and failure modes to be reviewed rather than treated as invisible model behavior.
Public source code