This paper proposes a novel control scheme, named self-reflective model predictive control (MPC), which takes its own limitations in the presence of process noise and measurement errors into account. In contrast to existing output-feedback MPC and persistently exciting MPC controllers, the proposed self-reflective MPC controller not only propagates a matrix-valued state forward in time in order to predict the variance of future state estimates, but it also propagates a matrix-valued adjoint state backward in time. This adjoint state is used by the controller to compute and minimize a second order approximation of its own expected loss of control performance in the presence of random process noise and inexact state estimates. The properties of the proposed controller are illustrated with a small but nontrivial case study.