A fundamental hypothesis to understand learning and memory is Hebb’s idea of the cell assembly, which persistent network reverberation can be sustained in a “closed-loop” circuit formed through activity-dependent synaptic plasticity. To test this hypothesis, we used cultured neuronal networks as a model system to investigate the spatiotemporal dynamics of neuronal ensemble activity. In these networks, brief stimulus at a single cell could sometimes evoke reverberatory activity lasting for seconds. Following the initial activation, a reverberation episode consists of rhythmic events of network activation at 5 to 10 Hz interleaved by short silent periods. Furthermore, after a few rounds of network activation, subsequent reverberatory events exhibited conserved spatiotemporal patterns of neuronal activation, reminiscent of attractor states in classical neural network models. This process was accompanied by an overall synaptic potentiation, and required functional NMDA receptors. These results provide direct evidence that activity-dependent synaptic plasticity underlies the formation of prototypic Hebbian cell assembly, and demonstrate how interconnected neurons can acquire emergent network properties through activity-driven self-organization.