Neo4j Labs releases Relate 1.0, an embedded Rust graph database for agents

Neo4j has announced Relate, an open-source graph database built to live inside an AI agent’s process rather than on a server. The repository tagged v1.0.0 on 11 November under the Apache 2.0 licence, and the engine is written in Rust.

What it is

The README describes Relate as an embedded, in-memory property-graph engine. A graph is one .rl file, a directory, or memory only, made durable with an fsync’d write-ahead log and Parquet checkpoints. You query it in Cypher. The same engine ships as a relate CLI, a Python package (pip install relate-db), npm packages for Node.js and the browser through WebAssembly, and a Rust crate.

For agents, three pieces matter:

• An agent skill that teaches coding agents such as Claude Code to write Cypher for Relate, installed with one command from relatedb.ai.
• An MCP endpoint, part of an optional single-process server that also speaks Bolt (so Neo4j drivers and neo4j-browser connect) and Neo4j’s HTTP Query API v2.
• Vector search in the same query language: vector() properties, similarity functions, CREATE VECTOR INDEX and a SEARCH clause for top-K with pre-filtering, plus hybrid vector and full-text queries.

The site’s examples are agent memory (“what do you remember from our last session?”), codebase context (“which functions break if you change parse_config()?”) and recommendations.

Read before you deploy

• The README opens with a warning: this is an experimental Neo4j Labs project and not part of Neo4j’s supported product lineup.
• The limitations page calls Relate single-process, “closer to SQLite than to a clustered graph server”. One graph per statement, no cross-graph queries, and most database administration commands are unsupported.
• The server binds to loopback with a generated auth token by default. RELATE_AUTH=none turns auth off, and serve is plaintext only, so put a TLS-terminating proxy in front of it if it ever leaves localhost.
• We could not load Neo4j’s press release, so we are not repeating any performance figures from it. The repo’s own compliance history says no current Cypher conformance pass rate is published.

Why this matters

Agent memory usually ends up as a vector store plus ad hoc JSON. A graph adds the “how is this connected” question that GraphRAG is built on, and Relate puts it in-process with vectors in the same query. It joins other open agent-memory projects such as Tencent’s HY Memory 2. Treat it as a Labs tool for local and embedded use, keep the server on localhost, and keep an export path while the project is young.