For an optimization, the code will read the MOs from the previous step as the initial guess for the current step (unless for some reason it cannot reuse the previous MOs, e.g. the number of linearly dependent MOs changed from one step to the next). There is currently no option that can be set to change this behavior.