When you look at how top engineering teams actually build agent memory systems, a pattern emerges: There is a filesystem interface for what agents see and database storage for what persists. The debate was never “filesystem or database;” it was always both, in the right layers.
The idea that filesystems make good agent interfaces isn’t new. Dust.tt was projecting company data into synthetic filesystems in mid-2025. Letta’s memory benchmarks showed filesystem tools outperforming alternatives. LangChain’s context engineering work laid…
