I was just fiddling around when I thought of this:

1) Window time domain data
2) FFT to obtain tones, tossing out stuff < -60 dBc or so.
3) Use tones to feed HB simulation.
4) iFFT results for time data.
5) Slide window and correlate with previous window.  If it changes enough then re-peat process, else use previous results with adjusted phase.

Seems simple enough but maybe I'm missing something.  Is this the method used for ADS envelope simulation?  This would not simulate memory effects but may work for simple input/output, such as behavioural models do in system simulators.  Or am I just missing something?  Any thoughts?