Tropical ecosystems are under increasing pressure from land-use change and deforestation. Changes in tropical forest cover are expected to affect carbon and water cycling with important implications for climatic stability at global scales. A major roadblock for predicting how tropical deforestation affects climate is the lack of baseline conditions (i.e., prior to human disturbance) of forest-savanna dynamics. To address this limitation, we developed a long-term analysis of forest and savanna distribution across the Amazon-Cerrado transition of central Brazil. We used soil organic carbon isotope ratios as a proxy for changes in woody vegetation cover over time in response to fluctuations in precipitation inferred from speleothem oxygen and strontium stable isotope records. Based on stable isotope signatures and radiocarbon activity of organic matter in soil profiles, we quantified the magnitude and direction of changes in forest and savanna ecosystem cover. Using changes in tree cover measured in 83 different locations for forests and savannas, we developed interpolation maps to assess the coherence of regional changes in vegetation. Our analysis reveals a broad pattern of woody vegetation expansion into savannas and densification within forests and savannas for at least the past ~1,600 years. The rates of vegetation change varied significantly among sampling locations possibly due to variation in local environmental factors that constrain primary productivity. The few instances in which tree cover declined (7.7% of all sampled profiles) were associated with savannas under dry conditions. Our results suggest a regional increase in moisture and expansion of woody vegetation prior to modern deforestation, which could help inform conservation and management efforts for climate change mitigation. We discuss the possible mechanisms driving forest expansion and densification of savannas directly (i.e., increasing precipitation) and indirectly (e.g., decreasing disturbance) and suggest future research directions that have the potential to improve climate and ecosystem models.