Why would you chain multiple advisors together instead of putting all that logic in one place?

Each advisor handles one concern (e.g. one for memory, one for RAG retrieval, one for logging) so they can be composed and reordered per use case, and a new cross-cutting concern can be added as its own advisor without touching the existing ones — the same single-responsibility reasoning that applies to servlet filters or middleware chains.