Empowering Memory Assistance: An Episodic Memory-Based Framework for Personalized Recommendations

ICLR 2026 Conference Submission17434 Authors

Keywords: episodic memory, dynamic environments, lifelong learning, temporal graph networks, multimodal representation, sequence prediction, continual learning, human-AI collaboration, cognitive architectures, assistive robotics

Abstract: Artificial agents operating in dynamic environments require the ability to recall and contextualize past experiences to inform future behavior. Drawing inspiration from human episodic memory, we propose a cognitively grounded recommendation framework that models time-evolving personal experiences using a dynamic, multimodal memory architecture. Our system encodes temporally structured actions, places, and interactions into a hierarchical temporal graph network (TGN), enabling agents to disambiguate overlapping behavior patterns and anticipate future actions based on long-term experience. Unlike traditional recommendation or forecasting models that rely on static, task-specific patterns, our approach supports continual memory updates without retraining, and generalizes across varied activity sequences. Evaluated on a structured dataset derived from three years of egocentric recordings, our model significantly outperforms state-of-the-art baselines (e.g., AntGPT, DyRep, Palm) on next-activity prediction and sequence alignment metrics. This work introduces a scalable, cognitively inspired memory architecture with broad applications in lifelong learning, assistive robotics, and human-AI collaboration.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 17434