MobileMem: Learning from a Year of Mobile Experiences

Abstract: As foundation models continue to advance, artificial intelligence is evolving from generative intelligence toward agents capable of long-term user understanding, continuous learning, and autonomous evolution. Memory will become a fundamental capability of next-generation AI, following perception, generation, and action, while edge devices will serve as a critical carrier of personal long-term memory. Focusing on on-device memory, this white paper systematically examines key scientific challenges including long-horizon personal modeling, memory organization, and continual cognitive evolution. It proposes a personal memory layer for terminal operating systems and a knowledge-driven approach to memory synthesis. Building on these foundations, we develop MobileMem, a long-term, multimodal, and evolvable benchmark for on-device memory, advancing the field from “storing information” to “forming cognition.” The white paper further outlines a three-stage development roadmap encompassing edge-cloud collaboration, on-device autonomy, and trusted collaboration, and envisions the evolution from personal memory toward a long-term cognitive infrastructure spanning devices and physical spaces. It aims to provide a new technological framework and industrial perspective for the development of next-generation personalized intelligence and intelligent terminals.