Online shopping has developed to a stage where catalogs have become very large and diverse. Thus, it is a challenge to present relevant items to potential customers within a very few interactions. This is even more so when users have no defined shopping objectives but operate in an opportunistic mindset. This problem is often tackled by recommender systems. However, these systems rely on consistent user interaction patterns to predict items of interest. In contrast, we propose to adapt the classical information retrieval (IR) paradigm for the purpose of accessing catalog items in a context of un-predictable user interaction. Accordingly, we present a novel information access strategy based on the notion of interest rather than relevance. We detail the design of a scalable browsing system including learning capabilities joint with a limited-memory model. Our approach enables locating interesting items within a few steps while not requiring good quality descriptions. Our system allows customer to seamlessly change browsing objectives without having to start explicitly a new session. An evaluation of our approach based on both artificial and real-life datasets demonstrates its efficiency in learning and adaptation.