SurrealDB ads

Launch week wasn’t long enough. So next week, we ship even more. SurrealDB 4, SurrealFS, SurrealDB Earth, and more. Simplicity or scale? Transactions or speed? Yes. Stop choosing.

In the real world, you don't get to tell your agent what kind of question it's about to be asked. A benchmark shouldn't either. Today we launched SurrealDB Agent Memory. Memory benchmarks are full of choices that never make the headline: which model answers, which model judges, whether half-right counts, whether the system knows the question's category. For Agent Memory we published all of ours. 77.2% on BEAM at 1M, with the official judge prompt. 57.1% strict, every rubric item met in full. The category never reaches our system. Every place we depart from the paper is listed, including the ten settings we switched on that the product does not enable by default. None of the memory vendors we checked publishes a strict score, so we counted theirs from their own per-question files. Read the judge before you trust the number.

Production AI agents need two graphs, not one. Graph engineering has mostly meant modelling the agent: nodes doing the work, edges deciding what happens next. The context those agents run on should be a graph too, but it's usually still a pile of chunks in a vector store. Get that wrong, and it shows up in production sooner or later. The agent cites something outdated, surfaces something it shouldn't, or can't explain what it read to get there. I break down what a real context graph actually needs, and why bolting a graph database onto a stack of separate stores doesn't fix it. 🔗 https://sdb.li/4ynVaAt

Cobrainer built agentic graph RAG and a graph-based memory agent on SurrealDB, with graph, vector, and full-text from one engine.

Building AI on sensitive data? Cobrainer runs its agentic graph RAG and graph-based memory agent on SurrealDB Cloud, managed with EU data residency.

Cobrainer runs its graph-based memory agent on SurrealDB, with graph, vector, and full-text in one multi-model engine.