What's this about?
A paper studying the memory pattern production agents actually use: a directory tree of markdown files that the agent itself reads, writes, and reorganizes with ordinary file tools — examining how such memory should be organized, how it evolves, and whether it stays sustainable over time.
What exactly is "filesystem-based memory"?
Instead of a special database, the agent keeps what it learns as plain text files in folders — like a person keeping a well-organized notebook. Claude Code's memory directory works exactly this way. It's transparent (you can read your agent's memory) but raises real questions the paper tackles: files go stale, duplicate, and contradict each other as they accumulate.
Why it matters
This is one of the first systematic studies of the memory design that won in practice while research focused on bespoke vector stores — useful directly to anyone running long-lived coding agents.