A former vector database employee is watching the AI ecosystem suddenly rediscover the humble filesystem, and it might be bigger than most realize. LlamaIndex published "Files Are All You Need" while LangChain explored filesystem-based agent context. Oracle compared filesystems to databases for agent memory, and Dan Abramov wrote about building a social filesystem on the AT Protocol.

This shift isn't about databases being bad—it's about a different approach entirely. Coding agents like Claude Code are driving this trend, with Anthropic reportedly approaching profitability largely due to CLI tools that read and write files. The bottleneck has moved from model capability to context management. Context windows in LLMs are like whiteboards that get erased, while filesystems offer persistent storage through simple file writing.

Recent research from ETH Zürich found that repository-level context files can actually reduce agent task success rates by over 20%, suggesting the problem isn't filesystems themselves but how people use them. The industry is moving toward a "skills over features" model where NanoClaw uses portable markdown files instead of MCP servers. The file format is becoming the API—Agent Skills as an open standard adopted by Microsoft, OpenAI, and others. This represents interoperability without coordination, where two apps that read markdown can share context without formal standards or partnership agreements.