Summary
Kay Malcolm, a database product management leader with 20 years at Oracle, makes the case that the database — not the model or the SDK — is the true last line of defense for enterprise AI agents. The talk starts from a real problem her team encountered: AI tools like Codex were making individual engineers faster, but not the team, because agent context generated during development was never committed to a shared store. Git captures code but not intent, leaving cross-timezone teams to rediscover reasoning that was already done.
Malcolm frames an enterprise agent as five layers — model, tools, context window, memory, and guardrails — and argues that the memory and guardrails layers are systematically underbuilt in most current deployments. Drawing on her background as a DBA and her experience with Oracle’s multi-model database strategy (relational, unstructured, graph, vector, and in-memory/Redis), she argues that agents increasingly need simultaneous access to all of these data shapes, and that the database is the only component in the stack positioned to enforce consistent guardrails, handle retrieval at the right granularity, and survive agent restarts.
The talk includes a personal narrative arc from SQL tuning at Southern Company through the introduction of unstructured, graph, and vector databases, used to illustrate why a single converged database matters for agents that need to traverse all of those data types in a single reasoning loop.
📺 Source: AI Engineer · Published September 14, 2026
🏷️ Format: Opinion Editorial
