We introduce a framework to design in-memory decision tree machine-learning (ML) circuits using memristor crossbars. Decision trees (DTs) offer many advantages over neural networks, such as enhanced energy efficiency, interpretability, safety, privacy, and speed, along with reduced dependence on extensive training data. We propose an adaptive multivariate decision tree (AMDT) training algorithm, which constructs decision trees that incorporate both univariate and multivariate features, facilitating the creation of higher accuracy and energy-efficient crossbar designs compared to the state-of-the-art (SOTA). Our circuits are realized using pure memristor crossbars, requiring just one memristor per cell and no transistors while employing sneak-paths for flow-based in-memory computations. In comparison to the SOTA, our approach produces designs that are, on average, 4% more accurate and require 12.6% lower energy.