A recent paper on arXiv argues that personal-agent memory has largely focused on learning what users know, prefer, and want. However, the authors point out that knowing the user does not necessarily mean knowing what the user and the agent have established together. This distinction is central to their critique.
The abstract begins to describe an assistant's proposal, but the text cuts off. Still, the core claim is clear: current memory systems may miss the shared history and mutual agreements that shape ongoing interactions. The paper suggests that grounding awareness in this joint context is a missing piece for personal agents.
If the argument holds, it would push agent design beyond static user profiles toward a more relational model of memory—one that tracks what has been co-created between human and machine. The source is a single preprint, so no comparison with other work is possible here.