Filesystems are having a moment

A former vector database employee reflects on the AI ecosystem's rediscovery of filesystems as a persistence and interoperability layer for AI agents. The post covers why context windows fall short as memory, how files like CLAUDE.md and SKILL.md are becoming de facto APIs for agent context, and the fragmentation problem with competing file formats. It references an ETH Zürich study showing context files can hurt agent performance when overly verbose, and argues that the real opportunity is files as a universal, user-owned interoperability layer between AI tools — not a replacement for databases, but a complementary interface layer backed by whatever substrate fits the use case.