Harnesses are not a solved problem, yet. There are multiple kinds of memory that are user in them, conceptually. Working, Procedural, Semantic, etc. depending on harness design and architecture. Observe-modify is an integral part of harness development that continues to evolve and change, model to model.

Adoption of a system sans functional observability and mutation is of little utility. There's no reason to use a system that cannot be altered in design cycles (or larger model-release cycles), when memory systems aren't that complicated to begin with. RAG, Vector, simple KV, append-only buffers, are all in competition and used at different scales for good reasons.