Whole-heart MRA techniques typically target predetermined motion states, address cardiac and respiratory dynamics independently, and require either complex planning or computationally demanding reconstructions. In contrast, we developed a fast data-driven reconstruction algorithm with minimal physiological assumptions and compatibility with ungated free-running sequences. We propose a similarity-driven multi-dimensional binning algorithm (SIMBA) that clusters continuously acquired k-space data to find a motion-consistent subset for whole-heart MRA reconstruction. Free-running 3D radial data sets from 12 non-contrast-enhanced scans of healthy volunteers and six ferumoxytol-enhanced scans of pediatric cardiac patients were reconstructed with non-motion-suppressed regridding of all the acquired data ("All Data"), with SIMBA, and with a previously published free-running framework (FRF) that uses cardiac and respiratory self-gating and compressed sensing. Images were compared for blood-myocardium sharpness and contrast ratio, visibility of coronary artery ostia, and right coronary artery sharpness. Both the 20-second SIMBA reconstruction and FRF provided significantly higher blood-myocardium sharpness than All Data in both patients and volunteers (P < .05). The SIMBA reconstruction provided significantly sharper blood-myocardium interfaces than FRF in volunteers (P < .001) and higher blood-myocardium contrast ratio than All Data and FRF, both in volunteers and patients (P < .05). Significantly more ostia could be visualized with both SIMBA (31 of 36) and FRF (34 of 36) than with All Data (4 of 36) (P < .001). Inferior right coronary artery sharpness using SIMBA versus FRF was observed (volunteers: SIMBA 36.1 ± 8.1%, FRF 40.4 ± 8.9%; patients: SIMBA 35.9 ± 7.7%, FRF 40.3 ± 6.1%, P = not significant). The SIMBA technique enabled a fast, data-driven reconstruction of free-running whole-heart MRA with image quality superior to All Data and similar to the more time-consuming FRF reconstruction.