We extend the time-dependent variational principle to the setting of dissipative dynamics. This provides a locally optimal (in time) approximation to the dynamics of any Lindblad equation within a given variational manifold of mixed states. In contrast to the pure-state setting, there is no canonical information geometry for mixed states, and this leads to a family of possible trajectories—one for each information metric. We focus on the case of the operationally motivated family of monotone Riemannian metrics and show further that, in the particular case where the variational manifold is given by the set of fermionic Gaussian states, all of these possible trajectories coincide. We illustrate our results in the case of the Hubbard model subject to spin decoherence.