Keeping priors in streaming Bayesian learning