The embedding problem of extrapolated microphysics into the macroscopic description is considered, if Non-Markovian effects are macroscopically important. A classical heavy particle in contact with many classical oscillators is described by a retarded Langevin equation, which turns out to be essentially different from a retarded Fokker-Planck equation, in particular for long time tail memory. The solutions of the retarded Fokker-Planck equation may become negative, so violating a fundamental property of probability distributions. An asymmetry for the embedding procedure of dynamical variables as compared with that of ensembles is discussed, probably manifest in our ability to prepare an initial ensemble.