Title: LifeSide: Evaluating Agents as Enduring Digital Companions

Abstract: To function as effective lifelong digital companions, AI agents must synthesize cues across multiple sessions, continuously refine their user models, and navigate evolving privacy constraints. Current assessment frameworks fall short in this regard, typically isolating tests for memory retention and short-term empathy. We address this limitation by presenting \benchmark, a new evaluation framework focused on multi-session loops involving Memory, Emotion, and Environment. This benchmark treats users as persistent entities within layered worlds, utilizing multi-agent simulations to integrate environmental shifts into conversational contexts while maintaining the distinction between internal states and outward expressions. Through extensive testing involving 2,000 distinct personas and 111,000 tasks, our results assess capabilities in memory tracking, user comprehension, privacy management, and emotional support. The findings highlight a significant performance gap: despite excelling in standard memory benchmarks, current models struggle to maintain accurate user understanding or provide genuine companionship over extended periods.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC