How neurons in the ventral premotor cortex (VPC) integrate and process multimodal sensory information remains unclear. We recorded VPC neuronal activity in two monkeys performing a tactile or acoustic bimodal detection task (BDT). Single-cell analyses revealed purely tactile, purely acoustic, bimodal, and decision-mnemonic responses. Population analyses of the VPC network and a recurrent neural network trained on the BDT uncovered notably similar latent dynamics. In both networks, neural trajectories diverged sharply from rest along modality-specific directions during stimulus presentation, while remained at rest during stimulus-absent trials. Subsequently, trajectories rotated toward a mnemonic subspace, where perceptual decisions (acoustic, tactile, and absent) were maintained in three distinctly different stable attractors. Using low-dimensional modeling, we propose a universal mechanism for the transition from sensory processing to memory retention explaining both biological and artificial latent neural dynamics in a simple dynamical motif. These results show that VPC neurons encode bimodal information, can integrate competing sensory inputs, and maintain multimodal perceptual decisions.