Memory traces in reinforcement learning