Search papers, labs, and topics across Lattice.
This study critiques the conventional evaluation of memory systems in conversational LLMs, which relies on Direct QA accuracy, by contrasting it with user satisfaction over a four-month deployment involving 40 users and 1,872 sessions. The findings reveal that while Direct QA scores vary significantly (19.7% to 70.1%), they do not correlate with user satisfaction, indicating that these metrics assess different capabilities. By introducing MemUse, the authors demonstrate that Natural Integration of prior context into conversations is more closely linked to user satisfaction, highlighting a critical gap between elicited retrieval and conversational relevance.
User satisfaction in conversational AI hinges more on the natural integration of context than on traditional memory recall accuracy.
Memory systems for conversational LLMs are conventionally evaluated by direct, fact-seeking questions about prior dialogue (Direct QA): can the model recall fact X from a prior conversation? We tested whether higher Direct QA accuracy correlates with higher user satisfaction in a 4-month deployment (40 users, 1,872 sessions, 7 memory conditions). Existing-benchmark Direct QA varies from 19.7% to 70.1% across the 7 conditions, but satisfaction does not change. We hypothesize that existing benchmarks and user satisfaction are tracking different capabilities: benchmarks measure elicited retrieval (recall when asked), while conversation requires natural integration (detecting relevance and naturally weaving prior context into a response). To examine this, we introduce MemUse, a set of real user-cued memory moments drawn from the deployment, scored by an integration-aware judgment of the natural conversational response. Holding the model and context fixed, the same system that scores 78.8% on Direct QA references only 7.9% of those facts in conversation -- a 71-point gap. Within these moments, Natural Integration is associated with satisfaction, whereas Direct QA is not. We release the deployment corpus and MemUse together with all judgments and scoring prompts at https://github.com/ryuichi-sumida/memuse.
