RealCompanion Benchmark for Long-term Conversational Memory
October 2, 2026
RealCompanion introduces a benchmark consisting of ten real AI-human relationships containing 27,218 messages spanning up to 120 days. The dataset includes profiles, personas, and reasoning traces to evaluate how models infer user identity and utilize longitudinal context.
HOW THIS AFFECTS YOU
●
researcherYou can use this dataset to move beyond recency-biased evaluation toward true long-term memory reasoning.