I propose starting with one primary memory engine behind an application-owned interface. Separate authoritative state, source evidence, and derived memory logically; add another engine for a demonstrated capability or isolation need, not for each named memory type.