Graph neural networks (GNNs) have emerged as promising approaches for graph-based indoor localization. Recent studies have incorporated GNNs to capture the spatial features of sensed data and improve localization accuracy. However, existing methods often struggle in complex indoor environments due to noise and signal instability, leading to suboptimal performance. In this paper, we propose LightLoc, a simple yet effective spatial‒temporal framework for accurate and robust indoor localization. Specifically, LightLoc constructs a fingerprint tensor and leverages a bidirectional long short-term memory (Bi-LSTM) network to model the temporal dependencies of fingerprints. The model is designed for easy deployment via common smartphones, eliminating the need for additional hardware. Furthermore, a dual-step graph convolutional network (GCN) is introduced to extract spatial correlations from neighboring fingerprint nodes. Extensive real-world experiments demonstrate that LightLoc achieves superior localization accuracy and enhanced stability compared with existing approaches.