This paper proposes a novel spatial filter for biomagnetic source imaging. The proposed spatial filter is derived based on a modified version of the minimum-norm spatial filter, and is designed to have a performance close to that of the adaptive minimum-variance spatial filter. In this method, the theoretical form of the measurement covariance matrix is estimated as an updated gram matrix in a recursive procedure. The proposed method does not require a large number of time samples, since the sample covariance matrix is not used. It can even be applied to single time-sample data, and is robust to source correlation. We have validated the method’s effectiveness by our computer simulations as well as through experiments using auditory-evoked MEG data.