In conversational search, users frequently correct misunderstandings, revise their requests, and return to earlier topics. According to the abstract for MemRevise, this means part of the dialogue history no longer reflects what the user actually wants. The authors argue that existing benchmarks do not isolate such turns, making it difficult to evaluate systems on this specific challenge.
The paper proposes MemRevise, a self-revising memory mechanism for conversational search. While the abstract does not provide details on the architecture or experimental outcomes, the motivation is clear: search systems should actively revise stored context rather than treating all past turns as equally relevant.
This work highlights a gap in current evaluation practices and points toward more adaptive conversational search, where the system can discard or update outdated information as the user's intent evolves.