The study investigates the computational mechanisms underlying flexible multitasking in artificial recurrent neural networks (RNNs). The authors identify "dynamical motifs" as recurring patterns of neural activity that implement specific computations, such as attractors, decision boundaries, and rotations. These motifs are shared across different tasks, allowing the network to reconfigure for new computations. The research reveals that these motifs are implemented by clusters of units when the unit activation function is restricted to be positive. Lesions to these unit clusters result in modular performance deficits, indicating that the modular organization of dynamical motifs is crucial for task performance. The findings establish dynamical motifs as fundamental units of compositional computation, intermediate between neurons and networks, and suggest that they may guide questions about specialization and generalization in whole-brain studies.The study investigates the computational mechanisms underlying flexible multitasking in artificial recurrent neural networks (RNNs). The authors identify "dynamical motifs" as recurring patterns of neural activity that implement specific computations, such as attractors, decision boundaries, and rotations. These motifs are shared across different tasks, allowing the network to reconfigure for new computations. The research reveals that these motifs are implemented by clusters of units when the unit activation function is restricted to be positive. Lesions to these unit clusters result in modular performance deficits, indicating that the modular organization of dynamical motifs is crucial for task performance. The findings establish dynamical motifs as fundamental units of compositional computation, intermediate between neurons and networks, and suggest that they may guide questions about specialization and generalization in whole-brain studies.