The essay showed up in February without much fanfare. A former Weaviate employee published it on a personal blog: files, they argued, are the best persistence layer we have for AI agents. Not vector databases. Not elaborate orchestration platforms. Just files.
It landed at a moment when the rest of the industry seemed to be converging on the same conclusion independently. LlamaIndex published "Files Are All You Need." LangChain documented its own filesystem-based approach to context engineering. Oracle weighed in with a comparison of filesystems versus databases for agent memory. Jerry Liu, who runs LlamaIndex, made the case bluntly: instead of a single agent juggling hundreds of MCP tool integrations, the more practical setup is an agent with a filesystem and five to ten tools—file access, a code interpreter, web search. That's the whole stack.
Andrej Karpathy offered a complementary read on why Claude Code works: it runs locally, on your machine, with your data. He framed this as a deliberate design choice, one that contrasts with approaches that push orchestration into the cloud and away from where users' files actually live.
The underlying problem is memory. Context windows behave like a whiteboard—information persists for the duration of a session, then it's gone. That's not how productive work tools function, and as coding agents have become the dominant real-world deployment of AI, the gap between what agents can do and what they can remember has turned into a practical bottleneck. Anthropic's Claude Code—a CLI tool that reads and writes local files—is said to be a meaningful contributor to the company's improving financial position, a data point that illustrates just how much traction the filesystem-native approach has gained.
The ecosystem's workarounds are multiplying. CLAUDE.md gives Claude Code project-specific instructions that survive across sessions. Cursor stores conversation history as searchable files. Developers have started passing around personal context files—sometimes called "aboutme.md"—that carry user preferences between agent applications without requiring any vendor coordination.
There's a complication, though. An ETH Zürich study found that repository-level context files—exactly the kind of CLAUDE.md setups that have become routine in coding projects—actually reduced task success rates across multiple agents and models. Inference costs climbed more than 20%. Agents given context files explored more broadly, ran more tests, and traversed more of the codebase. They were thorough. They were also slower to reach the actual problem. The files functioned like checklists agents took too literally.
The essay's author doesn't read this as a refutation. The design constraint it reveals is real, though: context files that work are minimal and precise. Verbose onboarding documents aren't context—they're documentation, and agents don't skim the way humans do.
The harder problem is fragmentation. CLAUDE.md, AGENTS.md, copilot-instructions.md, .cursorrules, and SKILL.md are all attempting to solve the same problem—giving agents durable, portable context—with no coordination between them. It's an early-stage format war. Anthropic's SKILL.md specification has attracted interest from other platforms, and frameworks like NanoClaw are experimenting with treating skills as composable markdown files rather than software modules, but the field remains unsettled.
The stakes of whichever format wins are not small. The standard that becomes canonical for agent context files will shape how AI agents store, inherit, and share knowledge going forward. Files are already how humans organize persistent information. If they become how agents do too, that's not a technical footnote—it's a structural change in what personal computing looks like.