Recurrence risks of cancer patient can change during treatment as a result of treatment-related tumor evolution. However, biomarkers that can monitor these changes are lacking. Here, we investigated whether tracking circulating tumor DNA (ctDNA) dynamics through liquid biopsy can inform real-time recurrence risk. Nasopharyngeal carcinoma (NPC) provides an ideal model where cell-free Epstein-Barr virus (EBV) DNA (cfEBV DNA), a ctDNA, can be sensitively detected. We conducted the EP-SEASON study (NCT03855020) and prospectively recruited 1,000 NPC patients undergoing per-protocol cfEBV DNA assessments at 11 time points and receiving sequential chemo-radiotherapy. Longitudinal cfEBV DNA displayed distinct patterns during neoadjuvant chemotherapy and radiotherapy. Despite the prognostic significance of cfEBV DNA at each time point, real-time recurrence risks changed in sync with cfEBV DNA dynamics. Furthermore, we identified phenotypes of whole-course ctDNA changing dynamics associated with different survival outcomes. In conclusion, tracking longitudinal on-treatment ctDNA can forecast real-time recurrence risk, facilitating risk-adapted, individualized patient management.