Shenglai Zeng, Xiao Yang, Tiffany Ma, Xinyuan Zhang, Nanshu Wang, Deren Lei et al.
arXiv 2610.22671 32 pages Read the paper
Personal memories can be relevant without being worth interrupting for; ProPerAgent links object-centered graph retrieval to a gate that refines actionable memories or abstains.
Proactive assistants can find a useful reminder yet still interrupt at the wrong time; ProPerMem tests whether one grounded intervention lands in the right window and silence holds elsewhere.
Proactive assistants can speak too early or repeat themselves; tool-integrated GRPO uses benchmark-verifiable rewards to train the generator to give one timely reminder—or stay silent.