<!-- Source: https://www.nitronedge.com/insights/neo4j-relate-open-source-embedded-graph-database-agents/ -->
# Neo4j launches Relate, an embedded graph DB for agents

Neo4j has released Relate, an open source in-memory graph database written in Rust for AI agents. What it does, why it matters and how to try it.

Published 2026-11-13

Neo4j has released Relate, a new open source graph database built for AI agents. It was announced on 11 November 2026, the day before the company's NODES 2026 online conference. Relate is written in Rust, runs in memory inside the application that uses it, and is published under the Apache 2.0 licence. Neo4j has also opened a hackathon to get developers building with it.

## What changed

Relate is a small, embedded database rather than a server. It runs as a library inside an application, needs no sign-up and starts in milliseconds. Graphs can live purely in memory, in a folder, or in a single file. An optional write-ahead log makes stored data durable, and state can be written to disk for long-term use.

It speaks Cypher, the query language used across Neo4j. Neo4j says Relate is fully openCypher compliant and more than 97% compliant with GQL, the ISO standard for graph queries. One query can combine vector similarity search with graph traversal, which is the basis of GraphRAG.

Several features are aimed squarely at agents:

- Errors come back in a structured form with a clear flag saying whether a retry might work, so an agent can recover without guessing.
- Results are capped at a set number of rows and marked when cut short, which keeps answers inside a model's context window.
- Transactions let several agents update the same graph without overwriting each other's work.
- It includes a built-in MCP server and a command line tool, and can detect common agent harnesses.
- A skill can be installed into coding agents so they find and use Relate on their own.

Existing tools can connect over the Bolt protocol or the HTTP Query API v2. The CLI installs from relatedb.ai, and Relate can be used as a library from Rust or Python, or in a web browser through WebAssembly. Neo4j says it also runs in Jupyter notebooks, containers and edge devices. Pricing and version numbers were not part of the announcement, and Neo4j has not said how Relate fits with AuraDB or its enterprise server.

Neo4j backed the launch with a Gartner forecast from June 2026: by 2029, machine users such as agents could account for 75% of database workloads, up from 10% in 2025.

## Why it matters

Agents need memory. They need somewhere to hold what they have learned in a session, how facts relate to each other, and what other agents have already done. Today many teams put that in a vector store alone, which finds similar text but does not record relationships. A graph keeps the links, for example which supplier serves which plant and which contract covers both. That lets an agent follow a chain of facts instead of guessing from loose matches.

The embedded design changes the cost of trying this. Each agent, user or sandbox can have its own small graph file, created and thrown away as needed, with no database server to run. That suits short-lived agent tasks and testing.

The agent-friendly errors and capped results sound minor, but they address real failure modes. Agents often loop on unclear errors or choke on huge result sets. Building those limits into the database removes work from every team.

Some caution is sensible. This is a first release, and teams should test performance, memory use and durability against their own workloads. An in-process database also inherits the security of the application it runs in, so access control, encryption at rest and audit logging need thought. Where agent memory holds personal data, someone must own how long it is kept and how it is deleted.

## What this means for organisations in the Middle East, India, Africa and North America

The local, embedded nature of Relate fits well with data residency rules. A graph that runs inside an application on an approved server, or on an edge device at a site, does not send data to an outside service. That is helpful for teams in the UAE and Saudi Arabia working under national data protection laws, and for Indian firms planning for the DPDP Act.

Edge support is relevant for oil and gas, utilities, mining and logistics operators across the Gulf and Africa, where connectivity at remote sites can be poor. An agent on a field device could hold a local graph of equipment and procedures and still work offline.

For banks and fintech firms in India, the Gulf and Canada, graphs are already a common tool for fraud detection because they expose links between accounts, devices and people. Relate could let an investigation agent build a temporary graph for each case. Any such use must still meet model risk expectations, such as OSFI guideline E-23 in Canada, which means logging what data an agent saw and why it reached its conclusion.

The Apache 2.0 licence also lowers the barrier for public sector bodies and universities that prefer open source.

## What to do next

1. Identify one agent use case where relationships matter, such as supplier risk, fraud links or equipment history.
2. Try Relate in a notebook or sandbox with a small, non-sensitive dataset and compare results against a vector-only approach.
3. Test the structured errors and row caps with your agent framework to see if they reduce failed runs.
4. Set rules for agent memory: what may be stored, where the files live, how they are encrypted and when they are deleted.
5. Decide how Relate would sit alongside any existing Neo4j or graph platform before moving beyond a pilot.

See our [Neo4j partner page](/solutions/data-analytics-governance/neo4j/) for more on graph databases, and our [agentic AI services](/services/agentic-ai/) if you are planning agent memory and data design.

## Sources
- [New In-Memory Database Purpose-Built for AI Agents](https://neo4j.com/press-releases/announcing-relate-by-neo4j/) (2026-11-11)
- [Announcing the Relate by Neo4j Hackathon 2026: Build AI Agents on an Embedded Graph Database](https://neo4j.com/blog/graph-database/announcing-the-relate-by-neo4j-hackathon-2026-build-ai-agents-on-an-embedded-graph-database/) (2026-11-11)
