It has been widely recognised that the difference between the level of abstraction of the formulation of a query (by example) and that of the desired result (usually an image with certain semantics) calls for the use of learning methods that try to bridge this gap. Cox et al. have proposed a Bayesian method to learn the user's preferences during each query. Cox et al.'s system, PicHunter, is designed for optimal performance when the user is searching for a fixed target image. The performance of the system was evaluated using target testing, which ranks systems according to the number of interaction steps required to find the target, leading to simple, easily reproducible experiments. There are some aspects of image retrieval, however, which are not captured by this measure. In particular, the possibility of query drift (i.e.~a moving target) is completely ignored. The algorithm proposed by Cox et al. does not cope well with a change of target at a late query stage, because it is assumed that user feedback is noisy, but consistent. In the case of a moving target, however, the feedback is noisy and inconsistent with earlier feedback. In this paper we propose an enhanced Bayesian scheme which selectively forgets inconsistent user feedback, thus enabling both the program and the user to ``change their minds''. The effectiveness of this scheme is demonstrated in moving target tests on a database of heterogeneous real-world images.