Rush DB
I’ve looked at RushDB, which bills itself as “instant memory for AI agents”: push any JSON, get a property graph, vector embeddings, and a queryable live ontology automatically. No schema design. No separate vector store. Backed by Neo4j under the hood. The envy of Neo4j really 🙂(or the thing to acquire, if you prefer).
There’s something genuinely clever here: the ontology API. Before an agent queries anything, it can fetch a compact snapshot of all labels, properties, value ranges, and relationships in the data. The agent knows the graph topology before it writes a single query, fewer hallucinated field names, fewer silent failures. That’s not a new idea in knowledge graph circles. But it’s new to most AI engineers duct-taping Redis + Pinecone + Postgres together for agent state.
A few honest observations here:
- Automatic graph means nodes linked by shared property co-occurrence. It’s not semantic relationship extraction. The graph emerges from structure, not meaning. Plenty of ammunition for RDF people who’ve never made peace with LPG.
- This is schema-on-read. Fast to start, hard to govern at scale.
- It’s a persistence and retrieval layer, not a knowledge graph tool.
- It’s early. RushDB 2.0 is recent, the free tier is generous (which means they’re still in growth mode), and the 150 stars on GitHub with 33 issues tells you that it’s something that has to prove itself.
Still, for agentic pipelines that need structured, searchable memory without a data engineering team behind them, it’s worth a look. The self-hosting story (Docker + your own Neo4j) is clean and it’s part of a growing ecosystem trying (with varying degrees of honesty and success) to bridge AI with KGs
Anyone here has moved this into production?