Persistent neural activity underlies key cognitive functions like short-term memory. To support persistent neural activity, circuits must learn to balance recurrent amplification (prolonging network time constants) with sufficient robustness to gate out perturbations in local parts of the network. It remains unclear how such persistent and robust representations emerge in neural circuits. Neurons in the mouse anterior lateral motor cortex (ALM) exhibit preparatory activity during motor planning, a robust persistent activity maintained by both hemispheres. We used longitudinal two-photon calcium imaging to track the same neuronal population in one hemisphere of ALM as mice learned a short-term memory task. Concurrently, we used optogenetics to perturb the other ALM hemisphere to probe network robustness over learning. Mice progressed from a naive, to learning, to expert stage over weeks, measured by the proportion of correct trials. Across learning, task selective neurons gradually emerged during all phases of the behavior, but particularly during short-term memory. At the population level, a persistent activity mode encoding upcoming choice (choice mode) emerged over learning, while corresponding action and stimulus modes remained relatively stable. Choice mode activity was initially sensitive to transient optogenetic perturbations, but became progressively robust, reflected by a higher retention of selective activity after perturbations. The choice mode was highly variable trial-to-trial in the learning stage, but became consistent by the expert stage. Using a topic model, we discovered that different "clusters" of neurons were responsible for choice mode activity on different trials in the learning stage, whereas in the expert stage, choice selective neurons were recruited more consistently across trials. These results outline a learning process where frontal cortical circuits first establish single-area persistent neural representations through recurrent amplification, followed by multi-regional integration and continual network refinement that increases the robustness of these representations.