Improving Computational Efficiency in Visual Reinforcement Learning via Stored Embeddings