Introducing MobileMem: Learning from a Year of Mobile Experiences 📱🧠 Paper: huggingface.co/papers/2608.1… Code: github.com/zjunlp/MobileMem Dataset: huggingface.co/datasets/zjun… Website: zjunlp.github.io/MobileMem/ Long-term memory is essential for personalized AI agents. Yet most existing benchmarks reduce a user to a sequence of conversations. Real life is far more complex. Personal knowledge is scattered across calendars, notes, bills, to-do lists, screenshots, photos, and app interactions in the realistic mobile environment. It evolves over time, and critical evidence may be buried among millions of tokens of distractors. MobileMem brings memory evaluation closer to this reality. We introduce a benchmark for testing whether agents can organize, update, retrieve, and reason over year-scale, cross-application mobile experiences. ⚡ What MobileMem offers: 📱 Mobile Memory Evaluation A trajectory in MobileMem contains heterogeneous experiences across diverse applications. 🧭 Knowledge-Guided Experience Synthesis Our KEME framework transforms fragmented sessions into temporally coherent, continually evolving user trajectories through top-down planning and bottom-up experience updates. 📝 Text + Multimodal Tracks The text track uses a predefined protocol for apps to send records as structured text events, while the multimodal track models non-integrated apps where assistants learn from user-shared screenshots. 🧠 Comprehensive Memory Tasks The MobileMem benchmark suite includes diverse question types, such as multi-hop retrieval, temporal reasoning, knowledge updates, relationship and preference understanding, visual reasoning, abstention, and more. 🔍 What did we discover? • Better retrieval does not always mean better memory: stronger systems may retrieve weakly related evidence and become more confidently wrong on unanswerable questions. • Converting images into captions is not enough. Captions improve visual reasoning by 3.50%, but reduce performance on other questions by 4.43% because of memory interference. • Controlled experiments suggest that well-designed distractors can be more challenging than simply extending trajectory's length. MobileMem reveals that future memory systems must do more than store additional information. They must update outdated knowledge, distinguish evidence from interference, reason across time and modalities, and know when not to answer. #MobileMem #LLM #Agent #Memory #MultimodalAI #Benchmark #NLP