Users’ Expectations and Practices with Agent Memory

AI agents have the potential to provide long-term personalized assistance to users, and this relies on effective long-term memory. While memory in agents has been extensively covered by prior work, there is little understanding of users’ expectations and practices with agent memory. As a preliminary investigation, we interviewed people who use AI tools with memory and analyzed online discussion posts of people’s experiences with such tools. We found that users often have incomplete mental models of how agents remember and recall information, and how their memories affect their behaviours. Users generally consider agents’ memories as belonging to different categories along a hierarchy from more generalized knowledge to more specific knowledge about the user or task. Users often desire the system’s memories to be cleanly organized by these categories. These findings reveal opportunities to design agent memory mechanisms to organize and control access to memories based on users’ task-based needs.