In this circuit,
Currently, the PNP and NPN driving the gain node are BFQ221/241 (high voltage high frequency complements), possibly resulting in lower leakage. But this change had no apparent effect on performance; leakage seems to be dominated by the capacitor alone (currently polystyrene), or the construction method, or external spookiness.
What's weird is, although it does appear to drift slowly, it is punctuated by infrequent (tens of minutes between events?) upsets, which cause it to drift several milivolts at a time. Which is surprisingly undesirable in this application.
What causes this? Is it cosmic rays? Can that induce enough charge (~10pC?) to account for the observation? Is it a material property of the capacitor or transistors?
Related question: what's the ultimate performance envelope of an analog S&H, and what is it limited by -- junction leakage, MOS leakage, radiation? What's the longest time constant that is theoretically / physically attainable under standard conditions (STP and background radiation)?
Tim