Filesystems aren’t replacing databases for AI agents—but they’re making us rethink when databases are overkill.

The short piece linked below looks at a quiet but practical shift in how agents store knowledge, and it nudges readers to rethink assumptions about infrastructure. Read it here: https://x.com/tuanacelik/status/2012212183833403889

The article points out a simple observation, one that feels obvious after the fact: agents increasingly keep context, conversation history, and even *skills* as files. For example, Claude’s agent skills feature is literally just directories of markdown files that teach an agent new capabilities, which makes the approach feel, well, very human. It’s easy to share, version, and edit, and it avoids server overhead.

At the same time, the author is careful, and so should readers be. Databases still have clear strengths, especially for *structured queries*, complex joins, concurrent access, and strong consistency. The argument is not that filesystems replace databases, but that a well-organized filesystem, paired with semantic search, is often enough for many agent workflows. In other words, match the tool to the problem.

A bit of history helps. Early agent systems piled on infrastructure because teams wanted reliability and scale. Lately, folks are discovering that many day-to-day needs are solved by readable files, simple directories, and a search index. That lowers friction for experimentation and sharing, and it makes onboarding less painful.

Looking forward the sensible bet is hybrid setups, where files are primary for human-editable knowledge and databases are used when structure, performance, or multi-user locking becomes essential. Practically, that means lightweight repositories of markdown for day-to-day agent behavior, with optional syncing to vector stores or databases as projects grow.

The tone of the piece is pragmatic, a reminder that sometimes the simplest tools win. It’s a small shift in engineering taste, and one that could make agent development feel more accessible, collaborative, and maintainable.