RetroDeckRETRODECK

What Is RushDB, and Is It a Good Database for AI Agents?

RushDB is an open-source, Apache-2.0 database in one Rust binary that stores graph, vector and full-text data on the same records and keeps a live catalog of fields agents can read before querying. It is a good fit for agent memory and evolving data, but replication lacks automatic failover and sharding is still a preview.

What RushDB is

RushDB is an open-source database, Apache-2.0, that ships as a single Rust binary. Version 3 replaced the earlier Neo4j-based 2.x line, which the project keeps on a separate branch with security fixes until October 28, 2027; the 3.0.13 release was published on November 20, 2026. Its README describes it as graph, vector and analytics on the same records, with a live schema that agents read before they query.

You install it with a shell script or Homebrew and start it with rushdb up, which serves HTTP on localhost with data in a local folder. The same process also speaks the PostgreSQL wire protocol, gRPC and Bolt, and there is an MCP server plus Python and TypeScript SDKs. RushDB Cloud runs the same engine as a hosted service. Windows is not supported natively; the docs point to WSL 2 or Docker.

The data model

RushDB's design starts from one observation: agents query badly when they have to guess field names. Everything in the data model serves that.

Records, labels and emergent schema

You write plain JSON, CSV, SQL inserts or Mongo-style documents. Each object becomes a typed record with a label. Types are inferred per value, and nested objects become separate linked records, so a company with departments with employees becomes three labels joined by relationships. There is no CREATE TABLE and no migration step. When one field holds different types across records, the catalog lists each type rather than rejecting the write.

The live catalog

Every commit updates a catalog of labels, properties, types, value ranges and relationship shapes, with exact counts. The video's author calls the layout a Labeled Meta-Property Graph. For an agent, the point is that it can read the schema first, then resolve the words in a question to real fields, look up how a value is actually spelled, even with a typo, and only then write the query. The MCP server exposes each of those reads as a tool, including findLabels, resolveNames, getSchemaMarkdown and searchValues.

Search and query in one engine

Vector search, exact or HNSW, runs over vectors you supply; the self-hosted engine does not generate embeddings, though RushDB Cloud can. BM25 full-text and hybrid search sit beside it, filtered by the same graph predicates. Queries can be written as JSON, SQL over the Postgres protocol, Gremlin, read-only SQL/PGQ or Mongo-style documents, with openCypher, ISO GQL and Bolt marked as release candidates.

What agent memory and retrieval need from a database

Before judging any database for agents, list what the agent actually asks of it. Six requirements cover most systems.

Scoping: memories must belong to a user, an agent or a session, and recall must never leak across them. Provenance: every recalled fact should point back to where it came from, so a wrong memory can be traced and removed. Mixed retrieval: real questions combine an exact filter, such as this account or this version, with semantic similarity and keyword match. Discoverability: the agent must know what fields exist, or it will invent them. Safe concurrency: several agents writing shared state at once must not corrupt it. Least privilege: an agent that only needs to read should be unable to write.

Plain vector stores cover mixed retrieval partly and the rest poorly. A relational database covers concurrency and privilege well but forces a schema up front. Graph databases cover relationships and provenance but often need a separate index for vectors. The comparison of graph and vector databases goes deeper on that split.

How RushDB measures up

On paper, RushDB answers most of that list in one place. The README lists agent memory as records scoped by user, agent or session and linked to the turn that produced them, with ready-made memory providers for OpenClaw and Hermes. Hybrid search runs under graph filters, so retrieval can narrow to the right account before ranking. The catalog addresses discoverability directly. Transactions and a write-ahead log handle concurrent writers, and capability tokens let you hand an agent read-only access that the engine enforces.

A November 18 video from the AI Anytime channel builds a four-agent system with a Claude Sonnet 5.5 orchestrator and three Haiku 5.5 sub-agents that store eBay listings and Reddit sentiment in one RushDB graph. By the video's own measurement, a follow-up question answered from stored memory took 37 seconds instead of 147, with no re-scraping. That is a single demo, not a benchmark, but it shows the pattern that pays: store what an agent learned once, query it instead of regenerating it.

Limits and when not to use it

Read the status and limits pages before you commit. Replication is asynchronous, followers are read-only, and promotion is manual with no automatic failover. Raft consensus and storage sharding are previews, and the docs say multi-shard commits are not crash-atomic. The default limits page lists 10,000 operations and a 60 second lifetime per transaction, and 100,000 records per label for SQL queries. openCypher and GQL conformance suites do not yet pass in full.

That makes RushDB a strong fit for agent memory, structured operational data and prototypes where the schema is still moving, and a weaker fit for workloads that need high availability today or very large single tables queried through SQL. If you already run Postgres with pgvector and your schema is stable, the catalog is the main thing you would gain, and you can approximate it by exposing your schema to the agent as a tool.

Where document knowledge fits

RushDB is built for records: listings, accounts, tickets, memories. Much of what agents need is prose instead: docs, notes, runbooks and code, where the work is chunking, embedding and keeping private text private. That is a separate layer, and many systems run both. RDK covers the document side: it indexes plain folders and Obsidian vaults as encrypted private chunks that stay on your node, and agents query it through MCP, the CLI or the Partner API, receiving a few hundred tokens of context instead of whole files. On repeated or reference-heavy work, that cuts token spend 80 to 90 percent.

Frequently asked questions

What is RushDB?
RushDB is an open-source, Apache-2.0 database that runs as one Rust binary and stores graph, vector and full-text data on the same records. Every write updates a live catalog of labels, fields, types and values, which agents read before they query. It speaks SQL over the Postgres protocol, Gremlin, Mongo-style documents and JSON queries, and ships an MCP server.
Is RushDB a vector database?
It includes vector search, exact and HNSW, plus BM25 full-text and hybrid search, so you do not need a separate vector store. But it is primarily a graph of typed records. The self-hosted engine searches vectors you supply rather than generating them; RushDB Cloud can generate embeddings for you.
Is RushDB production-ready?
Partly. The JSON query language, SQL, Gremlin and Mongo-style documents are marked stable. openCypher, ISO GQL and Bolt are release candidates. Replication is asynchronous with manual promotion and no automatic failover, and Raft and sharding are previews. It suits agent memory and evolving data well; check the limits page before relying on it for high availability.
How is RushDB different from Neo4j?
RushDB 2.x was built on Neo4j, but version 3 is its own Rust engine. It infers schema from JSON writes instead of requiring you to model nodes up front, keeps a queryable catalog for agents, and puts vector and full-text search in the same engine. Neo4j has a far longer production record and mature clustering.
What should an AI agent's database provide?
Memories scoped by user, agent and session; provenance for every recalled fact; retrieval that combines exact filters with semantic and keyword search; a schema the agent can discover instead of guessing; transactions so parallel agents do not corrupt shared state; and read-only credentials for agents that only need to look things up.