Mirror effects --- simultaneous increases in recognition accuracy for old and new items in a given condition --- provide an important benchmark for memory models, but only if they arise from memory-related differences between conditions. We present a novel approach to distinguish between decision-related (type I) and memory-related (type II) mirror effects in simple yes/no recognition paradigms using event-related potentials (ERPs). We modified a directed forgetting procedure (item-method) to specify and test a relationship between encoding differences (as measured by study phase ERPs), mirror effects (as measured by behavioural data) and ERP retrieval set effects (as measured by test phase ERPs) from the perspective of a strength-based signal detection model of recognition memory. New words were once blocked with old words cued to-be-forgotten (forget retrieval context) and once with old words cued to-be-remembered (remember retrieval context), which produced a mirror effect. In the forget retrieval context, recognition accuracy decreased and ERPs for correctly identified new words were less negative-going in the $\overline{\mbox{N400}}$ time-window (300--500~ms). This ERP retrieval set effect was unrelated to response-criterion shifts between conditions and may instead reflect changes in retrieval orientation, implying a type II mirror effect. Our results suggest that combining behavioural and ERP data can arbitrate between different theoretical explanations of mirror effects and thus memory models.