AIAny
Icon for item

MobileMem: Learning from a Year of Mobile Experiences

Provides a year-scale multimodal benchmark and evaluation framework for on-device long-term memory in personal assistants, built from real mobile user trajectories. Tests memory construction, retrieval, updating, temporal reasoning, and implicit preference inference, and includes a knowledge-grounded synthesis pipeline to form coherent long-horizon trajectories.

Introduction

Realistic personal assistants must remember and adapt to a user's evolving experiences over months, but most memory benchmarks reduce personalization to short dialogs or isolated facts. MobileMem flips that assumption: it treats a year of mobile usage as the unit of experience, asking memory systems to aggregate fragmented signals across apps, modalities, and time so agents can answer long-horizon questions and update beliefs as users change.

Key Findings
  • Benchmark design: MobileMem supplies year-scale user trajectories in both text and multimodal (omni) splits, pairing chronologically ordered app events, conversations, and screenshots with long-horizon QA that require multi-hop retrieval and temporal integration. This emphasizes experience-level memory rather than single-shot retrieval.
  • Synthesis & realism: A knowledge-grounded experience synthesis pipeline (used to produce coherent lifelong trajectories from fragmented sessions) enables temporally consistent scenarios where attributes, habits, and preferences must be inferred from distributed traces across apps.
  • Evaluation protocol: Systems process sessions in chronological order, incrementally build or update memory, and answer retained questions from the final memory state—making failures traceable to what the memory retrieved vs. model reasoning.
  • Research implications: Experiments reveal gaps in current memory systems, especially in updating changing preferences and aggregating sparse cross-app evidence, highlighting memory design as the primary bottleneck for long-horizon personalization.
Who it's for and tradeoffs

Great fit if you research or build personalized assistants, retrieval/memory architectures for agents, or multimodal long-horizon benchmarks and need realistic mobile behavior traces and targeted probes (e.g., temporal QA, implicit preference inference). Look elsewhere if you need purely conversational benchmarks, very large-scale public user logs (privacy constraints limit raw real-world data release), or turnkey memory systems—the contribution is a benchmark and synthesis framework, not a drop-in production memory service.

Information

  • Websitearxiv.org
  • AuthorsXinle Deng, Yida Xue, Xiangyuan Ru, Haoming Xu, Shuofei Qiao, Mengru Wang, Yijun Chen, Buqiang Xu, Chen Jiang, Yuchen Eleanor Jiang …
  • Published date2026/08/11

More Items

Hugging Face

Pairs metric-scale 3D scans of real environments with egocentric task recordings, estimated body and hand motion, and hierarchical annotations for physical AI research. The first release contains 126 scenes; the dataset is licensed CC BY-NC 4.0.

Turns LLMs’ physical knowledge into video training data using simulator constraints, physical verification, and iterative correction. It also repairs physical errors in generated videos and releases 100K videos spanning 2,500 physical events.

Tests decision models as actors, trajectory judges, and reinforcement-learning teachers for long-horizon agents. It finds that a model can be a weak actor yet still improve policy learning by evaluating actions after seeing their consequences.