Associative learning entails lasting changes in multiple, distributed neural circuits in different areas of the brain. Understanding the nature of information stored in individual memory circuits and how they interact cooperatively and competitively to function as one network are key but highly challenging problems. In Drosophila mushroom body (MB), ~20 pairs of dopamine neurons (DANs) and mushroom body output neurons (MBON) together form compartmental units of associative learning. Each DAN cell type is selectively tuned to rewards or punishments and write and update memories with cell-type-specific dynamics. Activity of MBONs is thought to be integrated by downstream neurons to guide memory-based actions and provide feedback to DANs to instruct future learning. However, its circuit mechanisms are still enigmatic. The latest EM connectome data revealed ~400 types of interneurons that have at least 100 synaptic connections with DANs and MBONs. The machine learning based prediction of their neurotransmitter and our new collection of genetic driver lines to manipulate them enable us to study how diverse local plasticity dynamics in each MB compartment and neural circuits for cross-compartmental interactions together define learning rules of the fly brain. Here we will present neural mechanisms of second-order conditioning. In the second-order conditioning, reward prediction by the first-order memory drives a formation of the second-order memory in the absence of reward. We found that the teacher compartment that has slow rate of learning and extinction instructs multiple students compartments with faster learning rate and higher flexibility. We identified two types of cholinergic interneurons that are 1) under control of competing inhibitory and excitatory drives from multiple MBONs whose activity represent memories about reward or punishment, 2) acquire enhanced response to reward-predicting cues after formation of appetitive memory and disinhibition from glutamatergic MBONs, 3) drive robust upwind steering when activated, 4) send highest number of excitatory outputs to multiple DANs, and thereby 5) induce secondary synaptic plasticity in their target compartments. This hierarchical interaction between MB compartments with distinct memory dynamics explains transient nature of second-order memory as originally described by Pavlov and Resorla and observed across animal phyla. Our results reveal the origin of action-correlated activity in DANs and how memory subsystems with distinct dynamics concertedly functions to enable higher-order learning.