Memory effects are the reason a PA that looks clean on a CW sweep falls apart on a modulated signal, and the DPD capture window is where most bench mistakes happen. First, capture length: too short a window and the DPD never sees the thermal tail; the PA looks corrected on the first burst and drifts as it warms up. I start with a capture at least several times longer than the slowest thermal time constant I expect, then trim back only if the correction stays stable across repeated bursts. Second, crest factor: a high-PAPR test signal exercises the PA peaks, but if the crest factor of the capture does not match the real traffic, the DPD learns the wrong operating point. I always record the CCDF of the capture alongside the DPD coefficients, so a later regression can be traced to a signal change rather than a PA change. Third, trigger alignment: a capture that is not aligned to the same point in the waveform from run to run adds apparent memory that is really just jitter. Check alignment by overlaying two raw captures before trusting any DPD result. Fourth, verify in the time domain, not just with ACPR: a DPD that improves ACPR while the EVM stays flat has usually just moved distortion around rather than removing it. My rule of thumb: if the fix does not survive a temperature soak and a crest-factor change, it is not a fix, it is a curve fit. How do you size your DPD capture window — do you size it from the thermal behaviour, the signal statistics, or just from what the instrument defaults to?