Scholars have long theorized a role for reactivation and rehearsal in maintaining items in memory. Recent evidence suggests that reactivation during awake rest or sleep predicts subsequent recognition and cued recall performance. In a free recall task however, the relationship between covert reinstatement and recall may be more complex because the memory cues must be self-generated and retrieval depends strongly on what has previously been retrieved. Retrieved context theories suggest that covert retrievals during a delay interval should specifically predict the initiation of recall. To test this hypothesis, we analyzed electrophysiological recordings from 116 neurosurgical patients performing a categorized free-recall task. Lists comprised 12 words (four exemplars from each of three taxonomic categories, drawn from a set of 25 categories), and subjects freely recalled these words following an arithmetic distractor task – a time period during which covert reactivation may occur, and in turn support subsequent recall. We trained an encoding model, using regularized regression, to predict oscillatory neural signals during the study period from the semantic (word2vec) representation of the category associated with each word. We validated our model by demonstrating that, during the encoding period on heldout lists, neural activity during word presentation is more similar to the model's prediction for the word's associated category than categories of other words on the list. We then compared model predictions for each category to neural activity during the arithmentic distractor task as a measure of neural reinstatement. We find that distractor activity was more similar to the model predictions for the first recalled category than others on the list and more similar to predictions for categories on the list than those not on the list, confirming the prediction from retrieved context models that reactivation would predict recall initiation.