A Dual-Memory Architecture for Reinforcement Learning on Neuromorphic Platforms