The time-convolution as well as the time-convolutionless generalized master equations (GME) are rewritten in a simple form permitting an exact formal solution. Using this form, it is argued that possible approximations must be performed consistently on both sides of the GME in order to control deviations of the emerging approximate solution from the exact one. Arguments are given that after such consistent approximations, the initial-condition term can no longer in general be taken as proportional to (scrI-
scrD0)\ensuremath{\Vert}${\mathrm{\ensuremath{\rho}}}_{0}$) (with
scrD0 being the projector used in GME). These conclusions are illustrated for the asymptotic-time symmetry-breaking problem of the spin-boson model. Although standard approximate GME approaches do indicate a possible symmetry breaking, we argue that these results are strongly influenced by the inconsistency of the approximations used. To support this statement we prove the absence of any asymptotic-time symmetry breaking in the zero-bias spin-boson model with a partial site-occupation relaxation.
Memory functions in time-convolution Generalized Master Equations (GME) for probabilities of finding a general system (interacting by a general coupling with a true thermodynamic bath) in individual states are considered without resorting to any approximation. After taking the thermodynamic bath limit, time integrals from zero to infinite times of the memories are considered. It is argued that these integrals entering, e.g., the usual naive Markov approximation converting GME the Pauli master (PME) equations are exactly zero. This implies long-time tails of memories (unobtainable by perturbational expansions) and slower-than-exponential long-time asymptotics of relaxation.