Abstract
<jats:p>Intelligent behavior requires neural representations to change with new demands while preserving learned structure. The prefrontal cortex is central to this ability, but the mechanisms are unclear. Here, we tracked medial prefrontal neurons for months as mice learned an association task with successive rule switches. Learning progressively stabilized when individual neurons were active during a trial, but not what task variables they responded to. Neurons repeatedly gained, lost, or changed selectivity even after their activity profile had stabilized. We developed Sparse Tensor Component Analysis to show that, rather than reflecting random drift, dynamic single-neuron selectivity arose through rule-dependent recombination of a fixed set of task representations. Thus, learning established a stable temporal scaffold in the prefrontal cortex within which neurons participated flexibly in different representations.</jats:p>