LifeSide: Benchmarking Agents as Lifelong Digital Companions

A new benchmark called LifeSide evaluates whether AI models can function as lifelong digital companions, finding that systems excelling at current memory tests still fail to sustain accurate user understanding and companionship over extended periods, according to research submitted in 2026 [1]. The benchmark, detailed in a paper submitted to arXiv, addresses what its authors describe as a critical gap in existing evaluations, which test memory recall and short-term empathy in isolation [1]. Lifelong digital companions must integrate cross-session cues, continually update their understanding of users, and adapt to shifting privacy boundaries, the researchers state [1]. LifeSide models users as persistent worlds with layered profiles and event trajectories, using multi-agent simulation to project environmental dynamics into dialogue while preserving the gap between latent thoughts and observable expressions [1]. The evaluation spans 2,000 census-grounded personas and 111,000 tasks across memory tracking, user understanding, privacy control, and emotional companionship [1]. Each persona is built upon 24-to-36-month event trajectories and environmental dynamics [3]. The benchmark introduces partial observability by modeling the gap between internal thoughts and spoken words, capturing the natural asymmetry of incomplete user disclosures [3]. It also integrates environmental dynamics, recognizing that external conditions constantly shift the significance of various life events [3]. LifeSide evaluates user understanding through two mechanisms: implicit inference, which tests whether an agent can infer latent user traits and preferences from interaction cues, and temporal user modeling, which tests whether the agent can dynamically update prior user understanding when later interactions provide revised information [4]. The privacy control component follows the framework of Contextual Integrity, treating privacy as a dynamic boundary rather than a static permission [4]. The findings align with broader challenges identified in the field. A separate benchmark, KnowMe-Bench, argues that memory is a necessary substrate but not a sufficient definition of person understanding, noting that a system can store and retrieve facts yet still fail to infer stable principles or connect distant experiences to present reactions [5]. KnowMe-Bench reconstructs autobiographical narratives into chronological cognitive streams with explicit timestamps and locations, decomposing text into visual observations, auditory inputs, situational context, background knowledge, and inner monologue [5]. The LifeSide researchers conclude that even models that saturate current memory benchmarks fail to sustain accurate user understanding and true companionship over long horizons [1]. The paper was submitted in June 2026 and is hosted on arXiv under the Computation and Language category [1].