For past experiences to guide our actions we need to retrieve the relevant memories. Here, we used electroencephalography (EEG) to investigate how a memory is selected for retrieval from a large store of mnemonic traces, and to evaluate how selection operates during the retrieval cascade. We analyzed data from two studies in which people studied objects in picture or auditory word formats, and later recalled them using either written words (Experiment 1, N=28) or line drawings (Experiment 2, N=28) as retrieval cues. We used multivariate decoding to quantify the reinstatement of study phase neural patterns when people successfully identified items that had been studied in the format currently designated as targets, versus non-targeted items. Neural reinstatement emerged by 500 ms post-stimulus, as did the established left parietal event-related potential (ERP) signature of recollection. Reinstatement was also target-selective (greater for targets than non-targets) when test cues overlapped more with targets, as had previously been shown for the left parietal ERP (Moccia and Morcom, 2021). In contrast, when cues overlapped more with non-targets, neural reinstatement was non-selective or reversed, unlike the left parietal ERP. We also tested for goal-directed reinstatement prior to retrieval cues, hypothesized to drive selection. When words were cues there was strong evidence of this proactive reinstatement, but it was not detected when pictures were cues. Together, the data suggest that selection can act at multiple stages of recovery of a memory and depends on both external cues and goal-directed control.