Neo4j releases Relate, an open source in-memory database built for AI agents
Neo4j's new Relate database runs inside an agent's own process and talks to agents directly, a sign that databases are being redesigned around software users.
The short version
- Relate is open source under the Apache 2.0 license and built in Rust
- It runs in-process, in memory, with optional disk storage for durability
- It includes an MCP server, a command-line tool, and Bolt and HTTP query endpoints
- Errors carry retryable flags, and a row cap stops large results from flooding an agent
- Neo4j cites Gartner: machine users reach 75% of database workloads by 2029
Why it matters
If software agents become the main users of databases, the tools they rely on will be shaped around how agents behave, not how people work. Relate is one early example, though the claims so far come from the vendor's own announcement. Anyone building agents should know that a lightweight, disposable database is now a product category.
What Neo4j released
Neo4j, a company that makes graph databases, has released Relate. A graph database stores information as things and the connections between them, which suits questions like "who is linked to whom, and through what." As Cyprus Shipping News reported, Relate is open source, written in the Rust programming language, and aimed at AI agents and the systems built around them.
The headline claim is speed of setup. Neo4j says a coding agent, meaning an AI that writes and runs code on its own, can create a database and begin building an application in milliseconds. The announcement also says it installs quickly and runs anywhere.
One caveat up front: the article is based on a company release, distributed through Business Wire. The performance figures are the vendor's own, and the source gives no independent benchmarks or pricing details.
In memory and in process
Two terms explain the design. In-memory means the data lives in the computer's fast working memory instead of being fetched from disk each time. In-process embedded means the database runs inside the same program as the agent, so there is no separate database server to set up or connect to.
Neo4j says reads in memory take under a millisecond, which matters because agents tend to read and write constantly as they work through a task. Memory is lost when a machine restarts, so Relate offers protection. A write-ahead log, a running record of changes written to disk before they are applied, can be switched on for durable stores, and the database can also flush its state to disk for long-term storage.
The company's product chief, Sudhir Hasbe, called Relate "the ultimate agent-native database." He described it as a light in-memory store that can fall back on durable disk storage when needed. He also said it is meant for agent sandboxes, agent memory, and mobile or edge devices, where computing happens close to the user instead of in a data center.
Features aimed at agent habits
Most of the announcement is about small design choices that reflect how agents fail and recover. Relate detects when an agent harness, the software wrapper that runs an agent, is calling it. When something goes wrong, it returns structured errors with an explicit flag saying whether retrying makes sense. Neo4j says this lets an agent correct itself without custom code to interpret error text.
Another choice limits how much comes back. Relate caps the number of rows in a result and adds a marker showing the result was cut short. The point is to avoid overfilling an agent's context window, the limited amount of text a model can consider at once. Agents then receive results of a predictable size.
Connectivity comes built in. Relate has a Model Context Protocol (MCP) server, a common standard for letting agents use outside tools, plus a command-line tool. It also exposes Bolt and HTTP Query API v2 endpoints, so existing graph drivers, orchestration frameworks and developer tools should work with it.
Graph queries and retrieval
Relate supports Cypher, the query language Neo4j uses for graph data. A single Cypher statement can search by vector similarity, which finds items with similar meaning, and follow connections through the graph in the same step. Neo4j says this supports multi-hop reasoning, where an answer depends on following several links in a row, and GraphRAG, which feeds a model facts drawn from a graph. The company says this gives agents more accurate results and fewer hallucinations, meaning invented answers. That claim is Neo4j's and is not tested in the source.
The release says Relate is 100% openCypher compliant and more than 97% compliant with GQL, the standard graph query language. Those numbers suggest queries written for other graph systems may carry over, though the source does not say how that was measured. It is released under the Apache 2.0 license, a permissive open source license.
The bet behind the product
Neo4j backs the launch with a Gartner forecast from June 2026. According to the release, Gartner expects machine users to pass human and AI assistant usage by 2029, reaching 75% of database workloads, up from 10% in 2025. It describes agents as generating 10 to 100 times more interactions than humans, which drives up computing, storage and streaming demand.
That is the argument for a database built around agents. Agents may query far more often and in far greater numbers than people, so tools designed for human-paced use may not fit. Whether Relate fits better than a database a team already runs is a question the source does not answer.
What to watch
- Whether independent benchmarks confirm the millisecond start-up and sub-millisecond read claims
- Whether developers adopt Relate for agent memory, or stay with databases they already run
- Whether other database makers add agent-specific features such as retry flags and result caps
Try it
If you are building anything with agent memory, write down what happens to its data on restart, what it costs at your volume, and whether you can export it later.
Sources and links
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.
Listen to the AI Wire headlines
Listen to the AI WireListen to our daily AI shows every morning: on your commute, in the shower, wherever you are
Listen to the daily showsFirst reported by Cyprus Shipping News · Read the original. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
