Reinforcement Learning with Auxiliary Memory