Temporal knowledge graphs as agent memory
The field has agreed what agent memory is: a bi-temporal knowledge graph - entities and relationships, each fact tagged with when it was true and when you learned it, retrieved by meaning, keyword and traversal together. But somewhere we started treating that memory as an integration project - a graph store here, a vector index there, keyword search bolted on, kept in step by hand. An agent doesn't need a memory system. It needs a memory.
In this webinar, Tobie Morgan Hitchcock (Co-Founder & CEO, SurrealDB) reframes the question: not how many systems agent memory takes, but what it takes for an agent to have one coherent memory - one that answers what it knew about something, as of a moment, and why, in a single breath. Working from one concrete scenario - a CRM agent that's watched an account for two years - he shows what a whole memory makes trivial, and a fragmented one makes fragile.
In this session you'll learn
- Why agent memory is honestly a bi-temporal graph, and what queries that unlocks (as-of, relationship-aware, invalidation) that flat vector recall can't.
- The real cost of the assemble-it-yourself stack: three systems to keep in sync, and where the consistency bugs actually surface.
- How to serve hybrid retrieval (semantic + keyword + traversal) without a three-system join at query time.
- A concrete data model for facts with validity intervals, and how invalidation / supersession should work.
- A decision framework: when the multi-engine stack is worth it, and when one store wins on TCO and audit.
