Brain machine interfaces (BMIs) are an important technology for restoring mobility in patients with paralysis and a promising tool for understanding the neural correlates of learning. BMIs have been implemented in mice to study decoding and adaptation during volitional activation of single neurons - yet little is known about how learning restructures the broad networks involved. Here we used 2-photon (2P) imaging over eight days as mice learned to volitionally activate pairs of neurons in the motor cortex (MC) or hippocampus CA3 to receive water rewards. We developed an open source pipeline that integrates the tracking of single neurons over days with real-time neural-activity algorithms. We found that mice can learn to volitionally activate single neurons from both MC and CA3. In the vast majority of successful trials, only one of the two positive ensemble neurons was responsible for the reward. However, in both MC and CA3 the neuron driving the majority of the reward could adapt and change over the training period. At the network level, we also found a mix of adaptation strategies including positive ensemble cells increasing their correlation with the rest of their network, negative ensemble cells decreasing their correlation to the network, or overall decreases in non-ensemble cell correlations. Our findings identify striking similarities in single neuron and network changes supporting learning for the same BMI task in two different brain areas. Moreover, they reveal heterogeneous strategies supporting the BMI task and may guide the development of individual-specific BMIs in future research and human applications.