The traditional ensemble-based recommendation systems rely on user meta-data to determine the weight distribution of base recommenders in generating recommendations. However, this approach may not be feasible in real-world scenarios where user information is limited, such as in ad-hoc searching. To address this challenge, we propose AIRE, a user-agnostic ensemble model that learns the representation of base recommenders from the history of interactions and memorizes the patterns and characteristics within these representations. Our proposed model takes the advantage of few-shot learning and learns the base recommender representation through a prototype network. The results of experiments demonstrate the effectiveness of AIRE in outperforming single state-of-the-art session-based recommenders. AIRE also proves to be able to distinguish input sequences and base recommenders effectively. The superior performance of AIRE highlights the importance of building representations for base recommenders in ensemble-based recommendation systems for ad-hoc inquiries. This work sheds light on the potential of ensemble learning for improving session-based recommendations and opens up new avenues for future research in this field.