Working memory as a representational template for reinforcement learning

Working memory (WM) and reinforcement learning (RL) both influence decision-making, but how they interact to affect behaviour remains unclear. We assessed whether RL is influenced by the format of visual stimuli in WM, either feature-based or unified, object-based representations. In a pre-registered paradigm, participants learned stimulus-action combinations, mapping four stimuli onto two feature dimensions to one of two actions through probabilistic feedback. In parallel, participants retained the RL stimulus in WM and were asked to recall this stimulus after each trial. Crucially, the format of representation probed in WM was manipulated, with blocks encouraging either separate features or bound objects to be remembered. Incentivising a feature-based WM representation facilitated feature-based learning, shown by an improved choice strategy. This reveals a role of WM in providing sustained internal representations that are harnessed by RL, providing a framework by which these two cognitive processes cooperate.