MobileMem benchmark trains AI agents on year-long mobile memory

MobileMem is a benchmark and framework introduced by a 17-researcher team to push AI agents beyond isolated question-answering toward persistent, personal assistants. The core problem: existing memory benchmarks don't reflect real mobile settings, where experiences are heterogeneous, multimodal, evolving, and deeply personal. MobileMem addresses this by grounding its evaluation in a year-scale collection of real mobile experiences, then using a knowledge-grounded synthesis pipeline to generate coherent, temporally consistent long-horizon trajectories from user-app sessions. This design gives agents a realistic, continuous stream of usage data to learn from and reason over.

MobileMem provides complementary text and multimodal settings that test critical capabilities: multi-hop and temporal reasoning, knowledge updating, and implicit preference inference. Instead of testing isolated facts, it models experiences as whole trajectories, letting agents remember the past, understand the present, and adapt to the future. The framework thereby moves memory beyond simple information retrieval toward experiential intelligence—a key step for on-device assistants that must continuously learn from users' lives. It is available as a technical report with a project page, offering code, data, and demos for the research community to build on.