The data foundation for agentic AI
Agents fail at the foundation, not the model. This paper sets out the case for a single multi-model engine that holds context, memory, and the scale to run them, from one agent to millions, and the limits of that approach.
Key takeaways
Why the data foundation, not the model, is the constraint on agents in production, and why bigger models and longer context windows have not moved it.
What an agent needs from data: a read-think-write loop, transactional state, relationships as first-class objects, and changes pushed rather than polled.
Atomic context: one engine where document, graph, vector, time-series and relational data share one transaction, one permission model and one audit trail.
The scale envelope, from an embedded library and SurrealKV on disk to SurrealDS on object storage and SurrealDB Earth, with one database per agent that scales to zero.
Agent Memory, PostgreSQL and MySQL compatibility, how other approaches compare, and the trade-offs and limitations of consolidating on one engine.
What's inside
- Executive summary
- The 2026 reality: the foundation is the constraint
- What an agent needs from data
- Context, semantics, and knowledge
- Why the fragmented stack falls short
- The architectural argument: atomic context
- From embedded to object storage: the scale envelope
- Memory is the missing half: Agent Memory
- Familiar SQL: PostgreSQL and MySQL compatibility
- Other approaches
- Trade-offs and limitations
- What this enables, and where to start
- Conclusion
- Technical appendix: a minimal illustration
