This work aims at sketching over time the activity of the brain, providing dynamic maps of cortical activations triggered by a stimulus. Raw activations are computed from MEG or EEG using a distributed source model with equivalent current dipoles (ECD) lying on the folded cortical surface. Exploiting the natural graph structure of the cortical surface and the high time sampling of MEG and EEG recordings, neural currents reconstructions are used to compute, in a robust and efficient way, tracks of cortical activations over time. This problem is casted into a Markov Random Field optimization framework that can be optimized in a few seconds using graphcuts-based algorithms, which guarantee global optimality of the solution. A label is assigned to each vertex in space and time, indicating active vs. non-active condition. Such approach computes a minimum cut on a weighted graph related to a cost function, imposing spatio-temporal regularity constraints on the activations patterns. Nodes of the graph are indexed over space and time. Edges code for the local regularity. Time information is extracted from the amplitude of neural activations for overlapping time windows of 20ms duration. The data cost associated to each node describes how well its neural activation is similar to one temporal template. The possible templates are defined by the most powerful events in each time frame. The closest template is selected by minimizing the relative error between the time series of a given node and the templates. To favor coherent active regions, edge weights are tuned over time and space. The method is illustrated and validated with MEG data on the somatosensory cortex for a finger stimulation experiment. The ECD are estimated using minimum norm on a three nested spheres head model. Cortical activations are tracked over time in the primary somatosensory areas, as early as 25ms after stimulus onset, during the displacement of neural activity along the postcentral gyrus all the way to the secondary somatosensory cortex in BA5.