Glonce

MobileMem benchmarks on-device AI memory from a year of phone use

researchThe Glonce editorsHugging Face🔥81Glonce rating81Impact72Novelty68Relevance100Surprise50Players32Credibility82
MobileMem benchmarks on-device AI memory from a year of phone use

AI agents are moving beyond systems that answer isolated questions toward persistent personal assistants meant to understand, remember, and continuously learn from a user's experiences. That shift needs long-term memory: the ability to accumulate and reuse user-specific experience over time, not just retrieve facts on demand. The paper argues that existing benchmarks are inadequate for realistic mobile settings, where a person's experiences are heterogeneous, multimodal, evolving, and deeply personal, so testing memory the way current benchmarks do misses what mobile use actually looks like.

To address that gap, the authors introduce MobileMem, a benchmark and framework for studying on-device long-term memory grounded in a year-scale collection of mobile experiences. It uses a knowledge-grounded synthesis pipeline to construct coherent, temporally consistent long-horizon trajectories out of user-app sessions, rather than relying on isolated snapshots. On top of that, MobileMem offers complementary text and multimodal settings that cover multi-hop and temporal reasoning, knowledge updating, and implicit preference inference, so an agent can be tested on connecting events across time, revising what it knows as facts change, and picking up on preferences the user never stated outright. The stated aim is for agents built and evaluated on MobileMem to remember the past, understand the present, and adapt to the future.

The paper frames this as a shift in what memory is for: by modeling experiences rather than isolated facts, MobileMem moves memory beyond simple information retrieval toward what the authors call experiential intelligence for continuous personal learning.

Key facts

  • MobileMem is a benchmark and framework for studying on-device long-term memory in mobile AI agents, grounded in a year-scale collection of mobile experiences.
  • It uses a knowledge-grounded synthesis pipeline to build coherent, temporally consistent long-horizon trajectories from user-app sessions.
  • The benchmark provides complementary text and multimodal settings covering multi-hop and temporal reasoning, knowledge updating, and implicit preference inference.
  • It is designed so agents can remember the past, understand the present, and adapt to the future.
  • The framework treats memory as modeling experiences rather than isolated facts, positioning the work as experiential intelligence rather than plain information retrieval.

Why it matters

Persistent personal AI assistants need memory that accumulates and reuses a user's experience over time, not just a retrieval system that answers one question at a time. The paper's premise is that current memory benchmarks do not reflect real mobile use, where experiences are heterogeneous, multimodal, evolving, and deeply personal. MobileMem is built specifically to close that gap by testing memory against a full year-scale span of realistic mobile experience rather than short, isolated interactions.

Who it affects

The benchmark targets researchers and developers building on-device or mobile personal AI agents that need to carry memory across long stretches of use rather than treat each session as a blank slate. The source text does not name the authors or their institutions, so no individuals or organizations can be credited here beyond referring to them as the paper's authors.

How to use it

MobileMem is offered as a benchmark and framework rather than a shipped product: a knowledge-grounded synthesis pipeline for constructing long-horizon, temporally consistent trajectories from user-app sessions, plus text and multimodal test settings that probe multi-hop and temporal reasoning, knowledge updating, and implicit preference inference. That combination lets a team build an on-device memory system and then check whether it can connect events over time, update stale knowledge, and infer preferences the user never stated. The source gives no pricing, license, or release details.

How solid is it

The paper's stated method is grounded in a year-scale collection of real mobile experiences, synthesized into trajectories designed to stay coherent and temporally consistent rather than assembled as disconnected facts. That said, the source text does not report dataset size, the number of trajectories, sessions, or users covered, nor any benchmark results or evaluated models, so there is no way to judge from the available text how well any given agent actually performs on MobileMem.

Risks and caveats

No performance numbers, dataset scale, evaluated models, author names, institutions, or release date appear in the source text, which limits how much can be verified about the benchmark's actual difficulty or adoption. The trajectories are also produced through a synthesis pipeline rather than captured directly from real users, so how closely the constructed experiences match genuine year-long mobile behavior is not established in the text.

“By modeling experiences rather than isolated facts, MobileMem moves memory beyond information retrieval toward experiential intelligence for continuous personal learning.”

— the paper