Embedded in-memory graph database built for agent loops, with MCP and Cypher.
What it does
Relate is Neo4j's embedded graph database aimed at the working loop of an AI agent rather than at a server rack. It is written in Rust, loads in-process, and starts in milliseconds, so a coding agent can create a store, run a query, and throw the whole thing away inside a single task. Reads happen in memory; persistence is opt-in through a write-ahead log or an explicit flush to disk.
Two choices stand out for agent use. Structured errors carry a retryable flag, so a caller has something to branch on instead of parsing an error string. Large result sets hit a row cap and arrive with a marker saying rows were dropped, which keeps a careless query from flooding a context window.
Connectivity is broad for a young project: a bundled MCP server, a CLI, Bolt, and HTTP Query API v2, so existing graph drivers and orchestration frameworks can attach without a rewrite. One Cypher statement can mix vector similarity search with graph traversal, which is the GraphRAG pattern most agent stacks end up hand-building. The project is Apache 2.0 and claims full openCypher with most of GQL.
Source: https://relatedb.ai/