The multi-turn example is where agent memory diverges from document RAG, and it deserves more weight than the standard two-phase description gives it. In a conversation the retrieval unit is not a paragraph, it is a decision, and decisions are scattered across turns and frequently contradict each other.

A user who said "make it blue" in turn 3 and "actually green" in turn 11 will have both retrieved by similarity, with nothing in the embedding to indicate which one is current. Recency and supersession have to be explicit metadata rather than something you hope the ranker infers. Worth separating the two failure modes as well: document RAG breaks when evidence is missing, conversational memory breaks when stale evidence is present. Different problem, different fix.