LOW-DIMENSIONAL ATTRACTOR DYNAMICS UNDERLIE FLEXIBLE MEMORY AND ACTION IN MOUSE PFC

Complex behavior requires neural circuits to maintain task-specific memories while dynamically mapping stimuli to appropriate actions. However, how cortical dynamics reorganize to support such flexibility—especially during task switching and interference—remains unclear. Here, we developed a dual-task paradigm in mice that combined a delayed paired-association (DPA) task with an embedded olfactory Go/No-Go task. Mice rapidly learned the dual task, with early errors primarily on unpaired DPA trials, and exhibited asymmetric interference: Go trials disrupted performance more than No-Go trials. To uncover the neural mechanisms underlying this flexibility, we recorded prefrontal cortex (PFC) and projected responses onto low-dimensional subspaces encoding sample and choice. Early in learning, population activity conveyed both sample (A/B) and choice (lick/no-lick) information; as training progressed, sample representations remained stable while choice representations became more selective. Specifically, activity was maintained in the 'no-lick' subspace, only shifting toward the lick region when appropriate. This adjustment prevented premature actions, better linking behavior to task demands. To probe the circuit mechanisms involved, we sequentially trained low-rank recurrent neural networks (RNNs) on the same tasks. RNNs trained on DPA developed attractors on a circular slow manifolds. With learning memory attractors shifted downward in state space, away from the action-related half-plane. A geometric model predicts that optimal performance arises when memory attractors occupy this region—highlighting a trade-off: optimizing one task impairs the other. Consistently, targeted optogenetic manipulation of PFC during the delay shifted attractor geometry upward or downward, selectively enhancing one task at the expense of the other.