Real-world decisions often require combining different cognitive computations. Recent findings suggest that, on the one hand, these computations are distributed across the cortex. On the other hand, cortical microcircuits seem to implement them via (near) orthogonal population-activity patterns. But how are these local geometries combined at the level of cortex-wide dynamics? To answer this, we developed a new task for mice navigating in virtual reality (VR) to temporally disentangle different cognitive computations. Mice first accumulate evidence to determine the predominant color of a random bilaterally symmetric checkerboard pattern in the sample region, then hold it in short-term memory along a delay region, and finally choose the side arm whose color matches the predominant color of the sample. Using widefield Ca2+ imaging of excitatory neurons from across the dorsal cortex, we observed distinct large-scale activity patterns during evidence accumulation, short-term memory, and choice. Strikingly, we found that these patterns were near-orthogonal by using demixed principal component analysis to estimate a low-dimensional set of axes that best capture activity related to each task variable. Further, the short-term memory axis was well predicted by separately measured spontaneous activity timescales, suggesting that some but not all computations co-opt intrinsic properties of cortical circuits. Finally, to assess whether task variables were stably encoded across task epochs, we trained separate linear decoders for each position in the maze. We could decode predominant sample color during both sample and delay, but the coding axes across the two epochs were largely orthogonal both globally and within areas. Overall, our findings reconcile recent findings on microcircuit and large-scale cortical dynamics by showing that activity patterns supporting distinct cognitive computations are distributed but disentangled. This could both prevent interference and facilitate generalization across task conditions.