If the need is "remember facts and preferences across sessions for a single assistant use case," a provider's built-in memory tool does that out of the box, with no infrastructure to maintain and no vector database to keep in sync.
Once the knowledge base is large, shared across multiple products or teams, or needs retrieval logic specific to your data (custom ranking, access control per document), a built-in generic memory tool won't flex that far — that's genuinely a custom-build problem.
Custom RAG isn't just the initial build — it's ongoing maintenance: re-embedding when the model changes, monitoring retrieval quality, handling edge cases. For a small team, that ongoing cost is often the deciding factor, not the initial setup.