During the Umbria-Marche, central Italy seismic sequence a small-aperture (≈200 m), four-station array was operating in the Colfiorito plain, a few kilometers away from the epicenters of the ML 5.6 and 5.8 mainshocks of 26 September 1997. The array was deployed approximately 500 m from the eastern edge of the basin. We analyze the three-component seismograms of 12 aftershocks, in a magnitude range of 2.5 to 4.1. Amplitudes of the horizontal components are systematically higher than those of the vertical component, with an average horizontal-to-vertical spectral ratio of about 3 at 1 Hz. In this frequency band, earthquake-induced ground shaking is highly coherent across the array. A 1-sec running-window zero-lag cross-correlation algorithm is used to compute apparent velocity and backazimuth of coherent wave trains in the frequency band 0.5 to 2 Hz. Apparent velocity and backazimuth show a different behavior in the first part of the seismograms compared to the late coda. The largest amplitude waves, that is, S waves and early coda, are characterized by low apparent velocities, mostly between 400 and 1200 m/sec. This suggests that, near the rock edge, the most significant part of seismic energy propagates horizontally in the basin. Backazimuth of these low-frequency, coherent wavetrains never coincides with the array-to-source direction. The predominant backazimuth is peaked around N110°, corresponding to the nearest, steep outcrop of the basin edge. The observed 1-sec coherent wave trains are interpreted as locally generated surface waves that are persistently diffracted from the nearby basin edge as long as a significant level of seismic radiation is incident to the bedrock. When the bedrock excitation decreases a much larger variability of both apparent velocity and backazimuth is observed, suggesting that, in the coda, randomly scattered waves within the basin and late arrivals of deeper origin become more important. Multipathing from the source to the site as well as multipathing within the basin are therefore interpreted as the main causes of the observed long-duration, coherent low-frequency basin shaking.