Does retrieved memory quality depend on its functional role?

Work on conversational RAG has overwhelmingly optimized the mechanics of memory — structure, retrieval size, granularity — treating retrieved context as undifferentiated. This paper's move is to ask what kind of memory was retrieved, not just whether it was relevant. With a fine-grained taxonomy of conversational memory roles and a user-centric evaluation that simulates user perspectives (rather than the usual reference-based scoring that flattens preference nuance), it shows the type matters: clarifying memory raises factual accuracy and constraint awareness, making responses more correct and personalized, while irrelevant memory does not merely fail to help — it reduces topic relevance and degrades constraint awareness. Memory can be a net negative, not just a missed opportunity.

The structural claim is that conversational RAG performance is driven by retrieving the right functional types of memory, not by maximizing relevance scores over a uniform pool. This reframes retrieval as a curation-and-diversification problem: rank and select by role, not similarity alone. It complements Why do time-based queries fail in conversational retrieval systems? — that note locates failures in query type, this one locates them in memory type, and together they argue conversational retrieval needs structure on both ends. It also gives an evaluation-grounded reason for Can agents fail from weak memory control rather than missing knowledge?: indiscriminate retrieval injects irrelevant memory that erodes constraint focus, which is exactly the control failure that note describes.

The caveat is that a role taxonomy is only as good as the classifier that assigns roles at retrieval time, and the paper measures the effects of roles more than it delivers a deployable role-aware retriever — the practical gap is operationalizing memory-role classification online. The finding that irrelevant memory actively harms also pushes against the "more context is safer" instinct: because added memory can degrade rather than dilute, the safe default is not to retrieve more but to retrieve discriminately, which means role-aware filtering is a robustness requirement, not just a quality optimization.