Efficient contingent learning requires actions and outcomes to be closely associated in time. Extensive prior work indicates that interaction between amygdala, frontal cortex, and striatum is critical for this type of learning. Here, we conducted neurophysiology recordings and applied computational approaches to understand how activity in this circuit varies when choice and reward are separated in time, intentionally decreasing animals’ ability to associate the two together. We trained 3 rhesus macaques to perform a probabilistic reward-learning task. In each session, animals learn the probabilities of receiving a juice reward associated with 3 unique stimuli. Reward probabilities were systematically reversed over the course of a session. In addition, the trace interval – the time between choice and reward delivery – was also varied in a block-wise manner. Monkeys are less likely to choose the best stimulus both when trace intervals were longer and immediately following reversals. Using high-density semi-chronic drives, we collected the activity of nearly 2000 neurons in orbitofrontal cortex (OFC), ventrolateral prefrontal cortex (vlPFC), anterior cingulate cortex (ACC), anterior insula (INS), amygdala (AMG) and striatum (STR) while animals perform this task. At the level of single neurons, we found that (1) amygdala, OFC and vlPFC most strongly encoded chosen value compared to other areas, (2) nearly all areas encoded reward, and (3) striatum and insula encoded the trace length. To further disentangle these representations, we inferred neural dynamics using data-constrained recurrent neural network models to study the directed interaction currents between regions. We identified temporally- and directionally- specific bottom-up currents from amygdala to ACC encoding both reward and trace length. Taken together, these results highlight the role of amygdala input to frontal cortex during reward learning, particularly during credit assignment. They also provide functional connectivity motifs that could be used to provide biologically realistic connectivity constraints on recurrent neural network models.