Background: Cerebrospinal fluid (CSF) can be used as a type of liquid biopsy to detect brain tumors. We aimed to explore the genetic profiles of CSF-derived circulating tumor DNA (ctDNA) to predict intracranial tumor response and monitor mutational evolution during systemic treatment in non-small cell lung cancer (NSCLC) patients with brain metastases. Methods: We conducted aprospective study of 92 newly diagnosed NSCLC patients with brain metastases. Paired CSF and plasma samples were collected at baseline, 8 weeks after treatment initiation, and at disease progression. Primary extracranial tumor samples were available for 58 patients and all samples underwent targeted next generation sequencing of 425 cancer-related genes. Results: At baseline, the positive detection rates of ctDNA in CSF, plasma, and extracranial tumors were 63.7% (58/91), 91.1% (82/90), and 100% (58/58), respectively. A high level of heterogeneity was observed between paired CSF and plasma samples, while concordance in driver mutations was also observed. A higher number of unique copy number variations (CNVs) were detected in CSF ctDNA than in plasma. CtDNA-positivity in baseline CSF samples was associated with poor outcomes (HR = 2.565, P = 0.003). Moreover, patients with increased concentrations of ctDNA in CSF after 8 weeks of treatment had significantly shorter intracranial progression-free survivals (PFS) than patients with decreased concentrations of CSF ctDNA (6.13 months vs 13.27 months, HR = 3.92, P = 0.007). Increased concentrations of plasma ctDNA were associated with shorter extracranial PFS (6.13 months vs 11.57 months, HR = 2.626, P = 0.032). From clonal evolution analyses, the accumulation of subclonal mutations in CSF ctDNA was observed after 8 weeks of systemic treatment. The clonal mutations remained more than 80% in CSF after 8-weeks of treatment also predicted a shorter intracranial PFS (HR = 3.785, P = 0.039). Conclusions: CSF ctDNA revealed unique genetic profiles of brain metastases, and dynamic changes in CSF ctDNA could better predict intracranial tumor response and track clonal evolution during treatment in NSCLC patients with brain metastases.