Background: Circulating tumor DNA (ctDNA) is a promising biomarker for detecting molecular residual disease (MRD) and relapse after definitive treatment in multiple solid cancers. However, the significance of ctDNA is rarely clarified in locoregional gastric cancer (GC). Here, we conducted a prospective and observational study to evaluate the utility of ctDNA in predicting the recurrence risk of GC. Methods: From October 2016 to June 2019, 100 patients with stage II/III resectable GC were recruited in this study (NCT02887612). Primary tumors and plasma samples were collected perioperatively and after adjuvant chemotherapy (ACT). Somatic variants were captured via a targeted sequencing panel of 425 cancer-related genes. The plasma of patients was defined as ctDNA positive only if one or more variants detected in the plasma presented in at least 2% of the primary tumors. All the patients received curative-intent standard-of-care therapy. Results: Preoperative ctDNA was detectable in 38 (38.0%) patients but showed limited value for predicting recurrence. After surgery (median days, 4), the plasma of 25 (25.0%) patients were still ctDNA positive and they had higher recurrence risk than the non-positive patients (hazard ratio [HR], 2.74 (95% CI: 1.37–5.48); P = 0.003). Forty-one patients had evaluable plasma after ACT and 10 (24.4%) of them who were ctDNA positive had remarkably higher recurrence and death risk compared with ctDNA-negative patients (recurrence-free survival [RFS] HR = 14.99 (95% CI: 3.08–72.96); P < 0.001; overall survival HR, 11.88 (95% CI: 2.38–59.24); P < 0.001). In particular, post-ACT ctDNA achieved better predictive performance (sensitivity, 77.8%; specificity, 90.6%) than both postoperative ctDNA and post-treatment tumor biomarkers (i.e., CEA, CA199, and CA72-4). In all multivariate analyses, ctDNA positivity was an independent factor of RFS. Patients with ERBB4 mutation in their primary tumors had poorer RFS compared to those were ERBB4 wildtype (HR = 3.46 (95% CI: 1.32–9.03); P = 0.007). A comprehensive model incorporating ctDNA status for recurrence risk prediction showed a higher concordance index (0.78; 95%CI, 0.71–0.84) than the model without ctDNA status (0.71; 95%CI, 0.64–0.79; P = 0.009). Conclusions: Postoperative and post-ACT ctDNA was associated with MRD and high risk of relapse in patients with stage II/III GC and can be utilized to guide GC management in post-surgical settings.