Neo4j has released Relate, an embedded graph database written in Rust and licensed under Apache 2.0. It targets AI agents, sandboxes and edge devices.

Neo4j has released Relate, an open-source graph database designed for AI agents and agentic systems. According to the company, the database runs in-process, meaning directly inside an application without a separate database server, and holds data predominantly in memory. Neo4j states that AI coding agents can create a database instance and start building applications within milliseconds. Relate is available immediately at relatedb.ai.

Neo4j, a vendor of a graph intelligence platform, announced the product on 11 November 2026. Relate is aimed at agent sandboxes, agent memory and mobile and edge devices. According to Neo4j, it runs in Jupyter notebooks, web browsers and software containers.

The company bases the release on a Gartner forecast. In “Forecast Analysis: Agentic AI in DBMS, Worldwide, 2026” (June 2026), Gartner projects that machine users will account for 75 percent of database management system workloads by 2029, up from 10 percent in 2025. Early enterprise adoption centered on conversational assistants for retrieving and summarizing information, Gartner notes, while budgets and workloads are shifting toward autonomous systems that plan, execute and coordinate multistep tasks. AI agents generate 10 to 100 times as many interactions as human users, which increases concurrent access and raises compute, storage and streaming consumption. According to the forecast, this expands enterprise infrastructure and lifts overall spending in the DBMS market. Neo4j argues that this calls for systems that offer fast data access and lightweight deployment under high interaction volumes.

On performance, Neo4j reports sub-millisecond reads for data held in memory, intended to serve the high frequency of read and write operations in agentic workloads. A single query in the Cypher query language can combine a vector similarity search with graph traversals. Transactional integrity is handled by an in-memory engine; an optional write-ahead log secured by fsync is available for durable storage. Users can also write the database state to disk using built-in functions.

Several functions address how agents work in execution loops. According to Neo4j, Relate detects agent harnesses automatically, meaning the software environments that run and control agents, and returns structured error messages that include a flag showing whether an operation can be retried. Agents can thus handle failed operations without custom parsing logic. A row cap limits the number of records a query returns, and a marker in the response indicates when results were truncated. This is intended to keep agents from overloading their context window and to let them recognize incomplete results. Multi-hop reasoning and GraphRAG are meant to improve the accuracy of results and reduce hallucinations, the company says.

Relate is published under the Apache 2.0 open-source license and written in Rust. Neo4j states that it is fully compliant with openCypher and more than 97 percent compliant with GQL, the standardized graph query language. For connecting agents, the database includes a server for the Model Context Protocol (MCP) and a command-line interface. Bolt and HTTP Query API v2 endpoints are intended to allow existing graph drivers, orchestration frameworks and developer tools to be used.

Sudhir Hasbe, President of Technology and CPO at Neo4j, described Relate as a lightweight in-memory store that can persist data to disk when needed. He named in-process use in agent sandboxes, as agent memory and on mobile or edge devices as the intended scenarios, and pointed to agent connectivity via MCP and multi-hop reasoning through full Cypher support.

A separate Neo4j publication on AI trends for 2027, dated 4 November 2026, covers related topics, though it does not mention Relate. It cites a McKinsey study from August 2026 in which 80 percent of respondents report individual productivity gains from AI, while 37 percent see a positive EBIT contribution at company level. Neo4j also points to an “inference paradox”: agents iterate in loops and consume more compute, so total costs rise even as token prices fall. The company describes agent memory as a field in which agents learn from repeated runs and pass distilled experience on to other agents, provided access rights allow it. Graphs, according to Neo4j, can link such experiences to their sources and their temporal relevance, and clear rules for verification, correction and forgetting are required. It further notes that MCP has spread quickly and now sits under the Linux Foundation, and cites a Gartner estimate that around 40 percent of companies will downgrade or decommission AI agents by 2027 because of governance problems.

By Jakob Jung

Dr. Jakob Jung is Editor-in-Chief of Security Storage and Channel Germany. He has been working in IT journalism for more than 20 years. His career includes Computer Reseller News, Heise Resale, Informationweek, Techtarget (storage and data center) and ChannelBiz. He also freelances for numerous IT publications, including Computerwoche, Channelpartner, IT-Business, Storage-Insider and ZDnet. His main topics are channel, storage, security, data center, ERP and CRM.

Contact via Mail: jakob.jung@security-storage-und-channel-germany.de

Leave a Reply

Your email address will not be published. Required fields are marked *

WordPress Cookie Notice by Real Cookie Banner