Disclosure

Telling the user what the system used about them and what else it holds — the affordance that makes a correction possible.

Why it matters in a memory layer

Retrieval withholds by ranking. A document that scores low is absent, and "why wasn't this shown" answers with a number — nothing decided, so nothing can be reported. A memory layer holds a model of a person, and the difference between did not rank and chose not to say is the whole of what makes personalisation acceptable.

On this course's corpus the model-driven context costs 51 tokens and the tuned retrieval path also costs 51 — the same six memories, reached by different routes. What the model adds is the last line of its disclosure: "held, not used: residence." Retrieval withheld the address too; it simply scored low, and would surface on a differently worded question.

Report what was withheld, not only what was used. A disclosure listing two attributes invites the assumption that two are all you hold, and the first time that proves false, every earlier disclosure becomes a lie by omission.

Connections

Taught in: Personalization Without Creepiness

Used in: Cold Start and Shared Accounts

Do not confuse with: Memory Observability