Spike timing relative to theta phase has been shown to mediate spatial and non-spatial information processing in medial temporal lobe (MTL) regions including hippocampus [1,2,3]. A prominent theory suggests that neural activity at different phases of theta oscillations encodes the serial order of items within short-term memory [4]. Based on recordings of spiking activity and local field potentials (LFPs) of epilepsy patients performing a multi-item sequential memory task, we show phase-dependent spiking during the delay, with preferred phases that were related to stimulus position within the sequence. We also employ Recurrent Neural Network Models (RNNs) trained to perform an analogous task to study neural circuits underlying temporal order memory. Similar to our empirical data, we observe that RNNs contain highly stimulus selective neurons, develop network oscillations and also exhibit phase-dependent activity related to item position. Interestingly, for most units in both recordings and RNNs, the ordering of preferred phases did not reflect the serial order of previously shown items. Our study provides empirical support for spike-phase coding for temporal order memory in humans. Using RNNs, we find that qualitative similarities between neural recordings and network activity, including similar spike-phase dependence on position, emerge simply from task optimization. Thus, our approach provides a generative computational framework to investigate functional interactions between single unit activity and oscillations in neuronal networks in cognitive tasks.