Abstract
Retrieval-augmented generation solved the wrong half of the memory problem. It made relevant text available to a model; it didn't make the retrieved text accountable, and in the standard arrangement the retrieved passage is concatenated into a prompt where it's indistinguishable from instruction. This paper describes Mnemos, the memory extension of my convergence substrate, which takes the opposite position: a recalled item is a proposal, carrying its provenance, subject to the same promotion gate as any other proposal, and incapable of becoming a fact without passing it. Two consequences follow. The first is a security property that falls out of the architecture rather than being added to it: retrieved text can't acquire authority, so indirect prompt injection reaches the context as an attributed proposal that policy may refuse. The second is a retrieval result. Mnemos ranks vector similarity and BM25 separately and merges the rankings by reciprocal rank fusion rather than merging the scores, and I show why this isn't a tuning preference: the two scores live on incomparable scales, and rank fusion is invariant under any monotone rescaling of either, which score fusion isn't. I close on the tension I haven't resolved — an append-only substrate and a memory that must forget aren't obviously compatible, and I state the problem rather than claim it solved.