Deep reinforcement learning with time-scale invariant memory