The discussion of the deformation behavior of metallic glasses remains largely phenomenological because of their strong structural disorder. Whereas currently the most successful model is the shear-transformation-zone (STZ) model, its microscopic underpinning are controversial. Here we propose a crucial link between the phenomenological model and the atomistic reality. The key concept was discovered in our study of the potential energy landscape (PEL) of metallic glass. We found that the activation process from the initial PEL minimum to the saddle point (SP) and the relaxation process from the SP to the next PEL minimum are statistically decoupled. Thus the SP is linked to a large number of PEL minima; the one-dimensional picture of PEL often used is extremely misleading. This is because whilst the SP is well-defined only in the PEL without the kinetic energy, in the real system at the SP is highly dynamic. The dynamic fluctuations due to many-body phase interference makes the SP extremely degenerate, almost to the point it resembles liquid. Because it “melts” at the SP it loses the memory and forgets where it came from. This decoupling of the activation and relaxation processes justifies the mean-field approximation and supports the STZ model. We derived a scheme to predict the thermal evolution of the PEL energy which agrees well with the results of the molecular dynamics simulation. We also discuss the observation of the structural evolution of relaxation, rejuvenation and anelastic polarization for metallic glass under stress or thermos-mechanical creep by anisotropic pair-density function (PDF) analysis by high-energy x-ray diffraction.