Cloud Data Integration Storage & Data

Neo4j launches Relate, an in-memory graph database for AI agents

Neo4j has introduced Relate, an open-source, in-memory graph database built in Rust for AI agents and agentic systems. The company positions it as a lightweight database that can run in the same process as an agent, including in sandboxes, notebooks, browsers, containers and on edge devices. Relate is available under the Apache 2.0 license.

The product is designed to support agent workloads that need repeated, fast access to connected information. Neo4j says Relate offers sub-millisecond reads when operating in memory, while allowing data to be persisted to disk when required. Its design targets the execution loops of autonomous agents, which can issue frequent queries as they plan and carry out tasks.

Graph queries and agent integration

Relate supports openCypher, the graph query language associated with Neo4j. The company says the database is fully openCypher-compatible and implements more than 97 percent of the newer ISO GQL standard. Its query support lets applications combine graph traversals with vector-similarity search in a single Cypher statement, a capability Neo4j says can help agents retrieve relevant context across connected data.

Neo4j describes multihop reasoning and GraphRAG—retrieval-augmented generation using graph data—as central use cases. Graph traversals can expose relationships across multiple steps, while vector search can identify semantically similar information. The product is intended to give agents both kinds of retrieval in one system; the company says this can improve result accuracy and reduce hallucinations.

Relate includes a Model Context Protocol (MCP) server for agent connectivity and a command-line interface for direct execution. It also exposes Bolt and HTTP Query API v2 endpoints. Neo4j says these interfaces provide compatibility with existing graph drivers, orchestration frameworks and developer tools, allowing teams to connect the database through familiar mechanisms.

Designed for agent execution loops

The database includes features intended to make agent interactions more predictable. Relate can detect agent harnesses automatically and return structured errors with explicit flags indicating whether a request can be retried. That information is intended to let an agent handle an exception and attempt recovery without relying on application-specific error parsing.

For large retrievals, Relate applies a row limit and includes a truncation marker in the response. Neo4j says the mechanism is designed to keep returned payloads bounded and reduce the risk of filling an agent’s context window. The limit and marker make truncation visible to the consuming system rather than allowing an unexpectedly large result to pass through without indication.

Relate uses an in-memory engine and supports transactional integrity, according to Neo4j. For persistence, deployments can use an optional write-ahead log configured with fsync, or flush state to disk for longer-term storage. This provides a choice between keeping data in memory for low-latency access and adding disk durability where an application requires it.

Neo4j’s view of changing database workloads

Neo4j says Relate reflects a shift from conversational assistants that retrieve and summarize information toward autonomous systems able to plan and coordinate multistep work. Such systems may generate many more database interactions than a person using a conventional application, creating different demands for concurrency and response time.

In support of that outlook, Neo4j cites a Gartner report, Forecast Analysis: Agentic AI in DBMS, Worldwide, 2026, published in June 2026. The company quotes the report as projecting that machine users will account for 75 percent of DBMS workloads by 2029, compared with 10 percent in 2025. The projection is Gartner’s, as cited by Neo4j, rather than a measurement of current database usage.

Neo4j president of technology and chief product officer Sudhir Hasbe said Relate is intended to be a small, portable store for agents, with optional durable storage and support for running close to where agent workloads execute. The company also says coding agents can create a Relate database and begin building applications in milliseconds. Those are vendor descriptions of the product’s design and behavior; performance in a particular deployment will depend on its workload and configuration.

Position in the agent database landscape

Relate enters a field that includes other in-memory and embedded databases used for agent memory, graph retrieval and related workloads. The source material identifies ArcadeDB’s embedded mode, Memgraph and FalkorDB as graph-oriented alternatives, and Redis or Valkey as options often used for short-lived state, conversation buffers and caching. These systems differ in data models and intended roles, so their presence does not imply direct feature equivalence with Relate.

Neo4j’s distinction for Relate is its combination of an embedded, in-memory graph engine, openCypher queries, vector search and agent-facing interfaces. Its portability may suit applications that need a graph store alongside an agent rather than a separately operated database service. The product’s persistence options, APIs and language compatibility are also relevant to teams evaluating how it might fit existing graph-based applications.

With Relate, Neo4j is extending its graph technology toward systems in which software agents, rather than people alone, initiate database work. The practical test for adopters will be whether the product’s query capabilities, integration paths and durability controls meet the requirements of their specific agent workloads.