This work studies learning-aided indoor fingerprint positioning (FP) for multiple-input multiple-output (MIMO) orthogonal frequency-division multiplexing (OFDM) systems, by exploiting their channel response matrices (CRMs). The existing learning methods of CRM-based FP mostly suffer from two main limitations: i) they require retraining when CRM dimensions change, which severely limits their practicality since the number of antennas and subcarriers may vary dynamically during practical implementation; ii) they overlook the temporal correlations in CRMs along the user trajectory, failing to exploit this information to improve positioning accuracy. Motivated by these limitations, we propose a novel deep learning-assisted FP method, named adaptive spatio-temporal neural network (A-STNN). This method consists of two key components: i) an adaptive mechanism that transforms spatial-frequency domain CRMs (SFCRMs) of arbitrary dimensions into truncated angle-delay domain CRMs (T-ADCRMs) with a unified dimension, enabling the neural network to handle varying antenna and subcarrier configurations; and ii) an STNN that exploits attention mechanisms to jointly extract spatial and temporal features embedded in T-ADCRM sequences along the user’s trajectory, thus improving positioning accuracy. Upon examination of simulation and measurement datasets, A-STNN achieves a prominent improvement in positioning accuracy over existing FP methods, in addition to its unique generalization capability across different antenna and subcarrier settings.