Learning is a process in which decisions and actions are adapted to satisfy the needs of an organism. In this talk I will discuss work in which we studied the neurophysiological mechanisms underlying learning from gains and losses. Monkeys were trained on a tokens-based reinforcement learning task. In the task, the monkeys learned to make choices that led to gains in tokens and avoid choices that led to losses of tokens. The tokens were periodically cashed in for juice rewards. While the monkeys carried out the task, we recorded neural activity from the orbitofrontal cortex (OFC), ventral striatum (VS), amygdala, and the medial, mediodorsal (MD) thalamus. We found that the monkeys learned well from the token outcomes. Analysis of single cells showed that the OFC maintained a strong representation of the number of accumulated tokens. Both the OFC and VS strongly represented the tokens received following choices in each trial, and the amygdala strongly represented the identity of the chosen stimulus. The MD thalamus, on the other hand, had an enhanced representation of the actions required to obtain the stimuli. When token outcome coding was examined in more detail, we found that the amygdala represented the salience of the outcome, responding more to large gains and losses, as opposed to monotonically encoding the value of the outcome. The OFC and VS both showed encoding of outcome value, particularly for gains. When we examined population coding, we were able to identify a subspace, strongest in OFC, that represented the token update, and the change in token number that followed token updates. When we examined this at the network level, using activity simultaneously recorded from all 4 areas, we found that token updates were calculated in the OFC and VS, by taking time-derivatives of the token count information. Overall, this suggests that the component operations that underlie performance of this task are implemented in a distributed network, composed of members of the classic limbic system.