In this paper, the robustness limitations of current recursive identification algorithms are highlighted. Then, it is shown how a particular non-recursive identification algorithm may improve the robustness to bounded disturbances and slowly time-varying parameters, at the expense of performing an on-line eigenvalue decomposition on a symmetric semidefinite positive matrix. Furthermore, contrary to "robustified" recursive identification algorithms, this algorithm does not require any a priori knowledge of a bound on the disturbances and of a bound on the unknown parameter values.