Measurements of neural activity in working memory during a somatosensory discrimination task show that the content of workingmemory is not only stimulus dependent but also strongly time varying. We present a biologically plausible neural model that reproducesthe wide variety of characteristic responses observed in those experiments. Central to our model is a heterogeneous ensemble of two-dimensional neurons that are hypothesized to simultaneously encode two distinct stimuli dimensions. We demonstrate that the spikingactivity of each neuron in the population can be understood as the result of a two-dimensional state space trajectory projected onto the tuning curve of the neuron. The wide variety of observed responses is thus a natural consequence of apopulation of neurons with a diverse set of preferred stimulus vectors and response functions in this two-dimensional space. In addition, we propose a taxonomy of networktopologies that will generate the two-dimensional trajectory necessary to exploit this population. We conclude by proposing someexperimental indicators to help distinguish among these possibilities.