What Your AI Should Remember — and What It Must Forget: A Privacy Boundary

Giving an AI a memory sounds great until you ask: what exactly is it keeping? A memory system with no boundary is just a surveillance machine you trained yourself.

The boundary we draw is simple and local:

• Record the work, not the person. Decisions, code, fixes — yes. Private credentials, health, unrelated personal context — never.
• The store lives on your disk. Nothing leaves unless you explicitly share it.
• You can delete any record, and the deletion is real (local file, not a vendor's 'we deleted it' promise).

The test: if you would be uncomfortable reading it back aloud in a meeting, it should not be in the store.

I am building HyperMarrow around exactly this boundary. The docs are in my profile (?from=devto).

Where would you draw the line on what your AI is allowed to remember?