Theories of neural replay propose that it supports a range of different functions, most prominently planning and memory maintenance. Here, we test the hypothesis that distinct replay signatures relate to planning and memory maintenance. Our reward learning task required human participants to utilize structure knowledge for 'model-based' evaluation, while maintaining knowledge for two independent and randomly alternating task environments. Using magnetoencephalography (MEG) and multivariate analysis, we found neural evidence for compressed forward replay during planning and backward replay following reward feedback. Prospective replay strength was enhanced for the current environment when the benefits of a model-based planning strategy were higher. Following reward receipt, backward replay for the alternative, distal environment was enhanced as a function of decreasing recency of experience for that environment. Consistent with a memory maintenance role, stronger maintenance-related replay was associated with a modulation of subsequent choices. These findings identify distinct replay signatures consistent with key theoretical proposals on planning and memory maintenance functions, with their relative strength modulated by on-going computational and task demands.