A long-standing theory proposes memories are stored in the synaptic wiring of strongly connected neural populations, known as neural assemblies. However, experimental evidence for assembly formation has been limited to inferring changes in synaptic connectivity from changes in neural activity following behavioral learning. Using tagged c-fos activity combined with channelrhodopsin expression, we directly measured the synaptic strength between neurons in the Piriform cortex (PCx) following mice learning an odor discrimination task. Our results suggest that an assembly forms after learning, but only for neurons tuned to the feature differentiating the odors, a process we call differential assembly formation (DAF). We explored potential neural circuit mechanisms for DAF by developing a computational model of PCx with biologically realistic synaptic plasticity mechanisms and both bottom-up stimulus and top-down reward-mediated inputs. Our model shows how disinhibition of pyramidal neurons (PNs) during reward promotes synaptic potentiation between PNs. By contrast, the absence of reward allows inhibition to depress synaptic learning between PNs tuned to non-salient features of the stimuli. The outcome is a strongly connected population tuned to the salient feature differentiating the stimuli. While inspired by findings from the PCx, our results are based on a disinhibitory circuit motif observed in multiple brain areas and found to be crucial for the ability to learn a broad range of behavioral tasks. Therefore, our findings may point to a general mechanism for learning that could exist throughout the brain.