Positive and negative subsequent memory effects have been reported in the left supramarginal gyrus (LSMG) which is thought to mediate attention to memory. We performed a secondary analysis of 142 verbal delayed free recall experiments, in patients with medically refractory epilepsy, undergoing pre-surgical evaluation with electrode contacts implanted in the LSMG. Convolutional neural networks (CNNs), utilizing high-gamma and beta bursts in the LSMG during the word encoding epoch, had an area under the receiver operating curve (AUROC) for classifying remembered words of >0.6 in 16 experiments. These experiments were distinguished by higher peak amplitudes of high-gamma and beta bursts and electrode placement in more anterior and dorsal regions of the LSMG. Among these 16 experiments the CNN utilizing high-gamma and beta bursts was equivalent, or outperformed, the equivalent CNN utilizing raw intracranial electroencephalogram (iEEG). In 7 of these 16 experiments, we also trained and tested CNNs that differentiated verbal word encoding from recall, and better performance recall from poor performance recall. These CNNs also utilized the high-gamma and beta bursts from the respective iEEG epochs. For many of the experiments these CNNs had an AUROC >0.6 demonstrating proof of principle that memory state and performance biomarkers can be derived from highgamma and beta bursts in the iEEG recordings of the LSMG alone.