MobileMem Tests Agent Memory Against a Year of Real Life
On Hugging Face's daily board at 20 upvotes: MobileMem (arXiv 2608.13606, 17 authors including Ningyu Zhang), a benchmark built from a year of real mobile user-app sessions, asking whether an agent can understand, remember, and keep learning from a person's accumulated experience. A knowledge-grounded synthesis pipeline turns the sessions into coherent, temporally consistent long-horizon trajectories, and the benchmark then tests multi-hop and temporal reasoning, knowledge updating, implicit preference inference, and on-device memory implementation. Project page: mobilemem.openkg.cn
The gap it targets is real. Existing memory benchmarks, LoCoMo, LongMemEval, test conversation recall, tidy dialogue histories where the answer is somewhere in the transcript. A year of phone usage is a different animal: heterogeneous, multimodal, personally relevant, and self-contradicting, because people change their minds, move cities, switch jobs. The knowledge-updating dimension is the hard one and the important one, a memory that cannot revise its beliefs is not memory, it is sediment. Implicit preference inference is the other quietly ambitious piece, what the agent should know about you is mostly things you never said.
The pairing with today's other memory story is the signal. On the builder side, ai-memory is trending on GitHub with compile-don't-retrieve markdown wikis for coding agents; on the research side, MobileMem is defining what passing looks like for life-scale memory. Builders and benchmarkers converging on the same layer in the same week usually means the layer is about to get good fast, that is how coding agents went.
arxiv.org/abs/2608.13606
← Back to all articles
The gap it targets is real. Existing memory benchmarks, LoCoMo, LongMemEval, test conversation recall, tidy dialogue histories where the answer is somewhere in the transcript. A year of phone usage is a different animal: heterogeneous, multimodal, personally relevant, and self-contradicting, because people change their minds, move cities, switch jobs. The knowledge-updating dimension is the hard one and the important one, a memory that cannot revise its beliefs is not memory, it is sediment. Implicit preference inference is the other quietly ambitious piece, what the agent should know about you is mostly things you never said.
The pairing with today's other memory story is the signal. On the builder side, ai-memory is trending on GitHub with compile-don't-retrieve markdown wikis for coding agents; on the research side, MobileMem is defining what passing looks like for life-scale memory. Builders and benchmarkers converging on the same layer in the same week usually means the layer is about to get good fast, that is how coding agents went.
arxiv.org/abs/2608.13606
Comments