the file system problem agents face

Databases are far from dead. Hot take within the vibe-coding community — but you can't build a reliable agentic memory system using files alone. The filesystem is a great interface for agents. LLMs are trained on repos, docs, and markdown. They're naturally good at grepping, reading file ranges, and writing artifacts. But there's a difference between interface and substrate. Files win as an interface. Databases win as a substrate. I ran a notebook comparing a filesystem-backed agent (FSAgent) with a database-backed agent (MemAgent) — same LLM, same task, same evaluation methodology. Three findings you need to see: → Small corpus + keyword queries: filesystem ≈ database. Both get there. → Large corpus + fuzzy queries: FSAgent scored 29.7%. MemAgent scored 87.1%. → Concurrent writes without locking: silent data corruption. Every time. Numbers don't lie. You need databases whenever any of these happen: 1. Concurrent writes from multiple agents or users 2. Semantic retrieval at sca