MobileMem: A Benchmark for On-Device Long-Term Memory in AI Assistants
A new framework called MobileMem has been introduced to study how AI agents can maintain long-term memory on devices. The details were shared in a paper on arXiv (ID: 2608.13606), which focuses on the challenge of building personal assistants that can remember and learn from users over time. MobileMem draws from a year’s worth of mobile interactions and uses a unique synthesis method to generate consistent and relevant long-term user trajectories. It includes both text and multimodal elements, addressing complex reasoning, knowledge updates, and understanding user preferences. This benchmark aims to improve upon existing ones that don't capture the rich and evolving nature of mobile user experiences. The paper was recently submitted to arXiv.
Key facts
- MobileMem is a benchmark and framework for on-device long-term memory.
- It is grounded in a year-scale collection of mobile experiences.
- It uses a knowledge-grounded synthesis pipeline.
- It constructs coherent and temporally consistent long-horizon trajectories from user-app sessions.
- It provides text and multimodal settings.
- It covers multi-hop and temporal reasoning, knowledge updating, and implicit preference inference.
- The paper is available on arXiv with ID 2608.13606.
- The announcement type is 'new'.
Entities
Institutions
- arXiv