A cold memory returns an error, and an error is not an answer

A customer writes: you promised me a refund last week. The support agent checks its memory before replying.

POST /agent/memory/search with the project key and the question. Recall is hybrid: BM25, vector, and graph fused: but two other properties matter more here. The question is privacy-stripped server-side before anything tokenizes or embeds it. And when the memory store isn't warm, the route fails closed: 503 store_unavailable, not an empty list.

That second property is the whole design. An empty result reads as "we never promised anything," a confident wrong answer sent to a customer. A 503 reads as "I can't check right now," which the agent turns into a hold and a ping to a human. Search that can't recall says so. It doesn't say no.

A cold memory returns an error, and an error is not an answer.