Files beat settings: Claude's memory is just markdown

ChatGPT's memory is a settings surface: a list you can review, entries you can delete, a "memory updated" chip when something gets captured. It's competent, discoverable, and completely opaque about how a memory will be used.

Claude Code went the other way: memory is files. CLAUDE.md for standing instructions, a memory directory for learned context. There is no memory UI because the file system already is one — visibility, editability, versioning, and provenance all come free from the medium.

The design lesson generalizes: when agent state is stored as a human-legible artifact, every interaction pattern around it gets cheaper. You don't build an inspector for what's already inspectable. The trade-off is ceremony — files demand more of the user than a chip that says "memory updated."

Both implementations are mapped part-by-part on the memory pattern page.