MobileMem: Learning from a Year of Mobile Experiences

Details

The next generation of AI assistants will need to accumulate user-specific context over months of interactions — not just answer isolated questions. MobileMem targets this capability with a benchmark grounded in a year-scale collection of real mobile app experiences, not synthetic Q&A pairs. A knowledge-grounded synthesis pipeline creates coherent, temporally consistent long-horizon trajectories from user-app sessions. The benchmark evaluates both text and multimodal settings across multi-hop and temporal reasoning, knowledge updating, and implicit preference inference — the full capability set a persistent personal assistant would require. This positions MobileMem against existing benchmarks that test isolated facts rather than continuous experiential context. The framing — remember the past, understand the present, adapt to the future — captures the gap well. For teams building AI companions, health apps, productivity assistants, or any personalization system, MobileMem offers a more realistic test surface than available alternatives. Published as a technical report with a public project page; not yet peer-reviewed.