Making a categorical decision about a stimulus can bias an observer's subsequent percept of the stimulus' parameters. We have previously demonstrated that these biases arise because human subjects seem to condition their percepts on their preceding decisions such that they maintain consistency across a sequential assessment of the same stimulus (Luu & Stocker VSS 2015). What remains unclear is how the conditioning is implemented. One possibility is that a decision directly modifies the stimulus representation in working memory such that the information that is inconsistent with the decision outcome is erased (implicit conditioning). Alternatively, the stimulus representation remains intact while the decision outcome is stored and used as an additional bit of information in the inference process downstream (explicit conditioning). We experimentally tested the two alternatives by probing how an incorrect categorical decision affects an observer's ability to reconstruct the stimulus parameter. Similar to our previous study, we briefly presented subjects with an array of small line segments whose orientations were sampled from a Gaussian centered around the true stimulus orientation. Subjects first had to report whether the overall array orientation was clockwise/counterclockwise relative to a reference orientation. After receiving feedback about their decision, they were asked to reconstruct the stimulus orientation by adjusting a pointer. Consistent across all five tested subjects, we found that in trials where subjects made an incorrect categorical decision their estimation biases depend on stimulus uncertainty. Implicit conditioning is not in agreement with this behavior. Because it assumes that the relevant stimulus information has been erased, it predicts that subjects' estimates are left to chance and thus are insensitive to stimulus uncertainty. In contrast, an observer model assuming explicit conditioning closely predicts the measured biases, suggesting that a categorical decision does not alter the stimulus representation in working memory in a sequential inference task.