Reviewed by MCPIndex
MCPIndex assessment
Memory MCP is a reference implementation for persistent agent state, not a production memory policy by itself. Its graph structure is refreshingly inspectable, but persistence makes provenance, identity resolution, confidence, and deletion mandatory. Use it for controlled continuity, never as an unquestioned fact store; a small graph of well-sourced observations beats a large transcript graveyard.
Quick overview
Memory MCP is not a chat history database; it is a small knowledge graph that lets an agent persist entities, relations, and observations beyond one context window. That distinction changes the failure model. A transcript stores what was said; a graph stores what the system believes about a person, project, preference, or dependency, which makes future retrieval compact and useful but turns every write into a potential long-lived falsehood. The server's core primitives — create entities, add observations, create relations, search nodes, and retrieve a graph — are deliberately simple, yet the semantic contract is not: the agent must distinguish fact, hypothesis, preference, and stale observation. The hidden edge case is entity collision. “MCPIndex,” “mcpindex.dev,” and “the directory project” may be the same entity, while two people with the same name may not be; careless creation fragments memory or merges identities incorrectly. Production memory needs namespaces, source and timestamp metadata, confidence, deletion or correction workflows, and a retrieval policy that never treats an old observation as current truth without checking its age. Persistence amplifies both continuity and error.