Communication is central to human life, yet it leaves humans vulnerable to misinformation and manipulation. Humans have therefore evolved a suite of psychological mechanisms for the evaluation of speakers and their messages. Here, we test a key hypothesized function of these "epistemic vigilance" mechanisms: the selective remembering of links between speakers and messages that are inconsistent with preexisting beliefs. Across four experiments, participants (N = 707) read stories associated with different contexts, with each story containing concepts that violate core knowledge intuitions ("counterintuitive concepts") and ordinary concepts. Experiment 1 revealed that after a brief distractor (2 min) participants more accurately attributed counterintuitive concepts to their speakers than ordinary concepts. Experiments 2a and 2b replicated this finding and found that this attribution accuracy advantage also extended to counterintuitive versus ordinary concepts associated with other contextual details-places and dates. Experiment 3 then tested whether this attribution accuracy advantage was more stable over time for speakers than for places. After a short distractor (20 min), there was a counterintuitive versus ordinary concept attribution accuracy advantage for both speakers and places. However, when participants were tested again after a long delay (48 hr), this attribution accuracy advantage more than doubled for speakers but disappeared entirely for places. We discuss the implications of these findings to the set of psychological mechanisms theorized to monitor and evaluate communication to guard our database of beliefs about the world. (PsycInfo Database Record (c) 2021 APA, all rights reserved).