Izzo: Okay, real talk — how many of you are running DuckDB, Postgres, Neo4j, and Pinecone just to get your AI agent to remember what happened yesterday?
Izzo: You're listening to Exploring Next, episode one-ninety-four. I'm Izzo, and with me is Boone. Today we're digging into SurrealDB 3.0 and their claim that they can replace your entire five-database RAG stack with one system.
Boone: And honestly? After looking at their architecture, I think they might actually pull this off. The transactional guarantees alone are fascinating.
Izzo: Right, because the current state of RAG is just... painful. You've got your vector database for embeddings, your graph database for relationships, Postgres for structured data, maybe Snowflake for analytics. And then you're writing orchestration code to sync all of this.
Boone: Exactly. And here's the thing that caught my attention — SurrealDB is storing agent memory as graph relationships directly in the database, not in application code or some external cache.
Izzo: Boone, break that down for me. What does 'agent memory as graph relationships' actually mean in practice?
Boone: So when your agent interacts with data, it's creating context graphs that link entities, decisions, domain knowledge — all as actual database records. An agent asking about a customer issue can traverse graph connections to past incidents, pull vector embeddings of similar cases, and join with structured customer data in one query.
Izzo: One query. That's the key difference here.
Boone: Right. Traditional RAG systems write separate queries for vector similarity, graph traversal, relational joins — then merge results in application code. Round-trips everywhere.
Izzo: And from a product perspective, that's where the accuracy problems come from. Your agent only gets partial context because each database only knows about its slice of the data.
Boone: SurrealDB's approach is storing data as binary-encoded documents with graph relationships embedded directly alongside them. SurrealQL — their query language — can traverse graphs, perform vector similarity searches, and join structured records without leaving the database.
Izzo: Okay, but here's what I want to know — how are they handling consistency at scale? Because if I'm running this across multiple nodes, I need guarantees.
Boone: This is where it gets really interesting, Izzo. Every node maintains transactional consistency, even at fifty-plus node scale. When an agent writes new context to node A, a query on node B immediately sees that update.
Izzo: No eventual consistency?
Boone: None. No caching, no read replicas. Everything is transactional. The CEO, Tobie Hitchcock, said their use cases involve data that's constantly updated where relationships and context need constant refreshing.
Izzo: That's... actually impressive. Most distributed systems make you choose between consistency and performance.
Boone: The Rust-native engine probably helps here. And they've got this Surrealism plugin system in 3.0 where the agent logic runs inside the database with transactional guarantees, not in middleware.
Izzo: From a go-to-market angle, they're targeting that sweet spot where you need multiple data types together. Not just vector search, not just analytics — but the full stack.
Boone: And they're honest about limitations. Hitchcock explicitly said if you only need analysis over petabytes of static data, go with object storage or a columnar database. If it's just vector search, stick with Pinecone.
Izzo: I respect that positioning. They're not trying to be everything to everyone.
Boone: The adoption numbers are solid too — 2.3 million downloads, thirty-one thousand GitHub stars. They're seeing deployments in edge devices, retail recommendation engines, ad serving tech.
Izzo: What about development velocity? Because that's usually the real sell.
Boone: Hitchcock claims what used to take months with multi-database orchestration can now launch in days. That's a bold claim, but if the architecture delivers...
Izzo: That's a B-plus from me. The technical approach is sound, the market timing is right with all the RAG complexity, and they're being realistic about their positioning.
Boone: I'm adding SurrealDB to the weekend project list. Want to see how that single-query magic actually feels in practice.
Izzo: Alright, build next time. Three things to go explore. First — install SurrealDB locally and try their SurrealQL syntax. See how vector search and graph queries look in the same statement.
Boone: Second, check out their Surrealism plugin system documentation. If you're building agents, you want to understand how to define memory patterns that run inside the database.
Izzo: And third — compare this to your current RAG setup. Map out how many databases you're currently hitting for a single agent query, then think about consolidating that into SurrealDB's unified model. The GitHub repo has solid examples, and their docs actually explain the architecture decisions, not just the API. That's a wrap on Exploring Next. Sometimes the best innovation is just... doing one thing really well instead of five things poorly.