The growing availability of Synthetic Aperture Radar (SAR) target datasets allows for the consolidation of different SAR Automatic Target Recognition (ATR) tasks using a foundational model powered by Self-Supervised Learning (SSL). SSL aims to derive supervision signals directly from the data, thereby minimizing the need for costly expert labeling and maximizing the use of the expanding sample pool in constructing a foundational model. This study investigates an effective SSL method for SAR ATR, which can pave the way for building the foundation model. The primary obstacles faced in SSL for SAR ATR are the scale problem of the remote sensing images and speckle noise in SAR images. To overcome these challenges, we present a novel approach called Knowledge-Guided Predictive Architecture (SAR-KPGA), which leverages local masked patches to predict the multi-scale SAR feature representations of unseen context. The key aspect of SAR-KPGA is integrating SAR domain features to ensure high-quality target features for SSL. Furthermore, we employ local masks and multi-scale features to accommodate the large image scale and target scale variations in remote sensing scenarios. By evaluating our framework on three target recognition datasets (vehicle, ship, and aircraft), we demonstrate its outperformance over other SSL methods and its effectiveness with increasing SAR data. This study showcases the potential of SSL for SAR target recognition across diverse targets, scenes, and sensors.