Personal Data

Memories carrying information about an identifiable person — including people who are not the account holder.

Why it matters in a memory layer

In a retrieval system PII is handled upstream: the corpus was assembled and redacted before indexing, and the read path inherits that decision. A memory layer manufactures its corpus one turn at a time, from the person the data is about, who is telling you on purpose. There is no upstream to defer to.

Which is why the obvious policy fails. On this course's corpus 7 of 37 memories carry personal data, and blocking all of it drops the gluten intolerance — health data, and one of the four facts the exam requires. Safety that removes the function is not a trade-off; it is one nobody measured.

Three of the seven are about the user's partner, detected by the entities link rather than by wording — "Sam is a nurse" and "I am a nurse" have the same shape and different consent stories. Those three are the easiest to justify blocking and the ones nobody proposes: they cost the user nothing and protect someone who is not in the room.

Connections

Taught in: PII on the Write Path

Used in: Deletion That Actually Deletes · Redaction and Minimization

Do not confuse with: The Memory Record