Using Language Models to Generate and Forget the Narrative Memories of an Assistive Robot

Angel F. Garcia Contreras, Wen-Yu Chang, Seiya Kawano, Yun-Nung Chen, Koichiro Yoshino

Abstract: We are developing Indy, a companion robot that aims to help people in their lives. This requires that Indy interacts and learns from experiences on their environment. This type of long-term cognition poses many challenges, for example, in deciding how and for how long the robot will “store” memories from multimodal observation for life-long learning. We have designed a memory system inspired by a cognitive framework, to help the robot remember important/relevant memories over long periods of time by forgetting irrelevant data. This includes heuristics inspired by cognitive psychology to determine when to “forget” information. In this demonstration, we show our first proof of concept for a “Narrative Memory,” a first-person narration of Indy’s actions and interactions. We use Large-Language Models (LLMs) to create an narrative episodes, and then progressively summarize them over time as a form of “forgetting”.

External IDs:dblp:conf/mmm/ContrerasCKCY25