You can apply the multilevel adaptive Gaussian quadrature algorithm of Pinheiro and Chao (2006) to multilevel models. This algorithm reduces the number of random effects over which the marginal likelihood needs to be integrated, significantly reducing the computational and memory requirements.