The "correction has to win at retrieval time rather than sitting alongside the original" line is the part I'd push on. That's easy to state and hard to build. Most vector stores and RAG pipelines don't have a native concept of supersession, if you correct a fact but don't explicitly delete or overwrite the old embedding, both versions get retrieved and the model is left guessing which one is current. In practice that means "correctable" isn't just a property of the memory, it's a property of your retrieval layer too, and it needs the same rigor as the promotion step you describe. Curious whether your org-tier memory versions entries or just overwrites in place, since those two approaches fail very differently when two agents read at slightly different times.