The question of how animals make decisions and adapt their behavior in a dynamic environment – known as behavioral flexibility – is crucial for survival. It requires a balance between exploiting acquired knowledge (exploitation) and exploring potentially better alternatives (exploration). Despite its significance, the intricate computational and neural mechanisms that underpin these processes remain unclear. To address this question, we designed a task in which head-fixed mice acquire and reverse self-initiated lever-pressing preferences in the absence of instructive external cues. We combined two-photon calcium imaging with behavioral modeling and optogenetics to investigate the role of the secondary motor cortex (M2), a higher-order structure involved in motor planning and spontaneous action initiation. Within the computational framework of reinforcement learning, we identified a neuronal signature in M2 as the main driver of history-dependent reinforced choice. Specifically, we found that M2 pyramidal neurons encode decision-value (DV) as persistent pattern of population activity, which serves as a short-term memory reservoir of DV that is critical for decision-making. Furthermore, the population activity related to DV exhibited orderly structured geometries in specific decision states, which facilitated consistent updates of DV between trials. To explore the mechanisms guiding transitions between these decision states, we monitored in M2 the axonal activity from thalamic nuclei known for their crucial roles in flexible behavior. In particular, we found specific clusters of axons from the mediodorsal and ventro-anterior/ventrolateral thalamic nuclei, that exhibited time-locked activity during key periods of the task, suggesting a distinctive contribution to the computational flexibility of M2 circuits.