A couple of days ago, two of my friends - independently - discussed projects involving relatively large but precise (low leakage, low temperature coefficient, etc.) capacitors. Both expressed concern about dielectric absorption (DA) being potentially detrimental for their applications.
DA is one of those pesky "secondary effects" which make our parts so different from the ideal ones. The most dramatic way how it demonstrates itself is when charged capacitors are discharged momentarily, and then without outside source they "regain" part of the voltage they were charged into. DA is caused by charges trapped in quantum states with relatively long release time - contrary to "elastic" dipoles which are behind the "classical" permeability of the dielectric and which act "instantaneously" with applied electric field, making up a "perfect" capacitor (up to a certain frequency of course). DA is modeled by several series RC, in parallel to the "root" capacitance (plus any "conventional" parasitics, e.g. the resistance representing leakage), with time constants in minutes to hours, and capacitances up to several percents of the "root" capacitance.
So, to my friends, I recommended foil capacitors off top of my head, but then remembering Bob Pease's articles on "how soakage in Teflon is way better than Mylar", immediately went to search for solid data. And they are not that easy to find. First of all, there are several types of foil capacitors with different dielectrics, which obviously implies different amount of DA - and long are gone Pease's times with dozens of different foil capacitors; today there are basically only three materials used: polypropylene (PP), a family of polyamides (polyesther terephtalate PET, basically your plastic bottle, that's Pease's Mylar; and polyesther naphthalate, PEN), and polyphenylene sulfide (PPS). There's no polystyrene anymore (probably because of its low melting point, making it unsuitable for modern SMD and RoHS processes), no polycarbonates (maybe the same reason), and no PTFE (Teflon) touted as the best of bests, at least not in the common commercial offerings - they were said to be notoriously hard to process, mainly metallize.
Okay, but how ceramics perform with regard to DA? They were only in nascent state in Pease's times - a plain ceramic disc with two metal foils glued on - and technology made giant leaps in this area in the past few years, with intense research both in dielectrics, and their processing into the MLCC chip capacitors which are commonplace today. Manufacturers understandably tend not to publish more than absolutely minimum data on their parts, and DA is rarely if at all specified. From what I've found, the absorption capacitance for C0G/NP0 ceramics were said to have be around 0.5%, whereas PPS foil was said to have absorption capacitance 0.02-0.05% of the nominal capacitance.
And then I stumbled upon aresearch on ceramic capacitors, performed for NASA by Alexander Teverovsky, and looking at the numerous graphs of current vs. time flowing into or out of a charged/discharged ceramic capacitor, in log-log scale - it finally downed to me:
(The second conclusion from Mr. Teverovsky's research is, that the absorption capacitance in X5R and similar high-permittivity capacitors is surprisingly high, tens of percents, typically around 25-30%; again, the difference to the much lower value in other sources - often around a few percents - can again be explained by the "unpatient" standard methodology of measuring DA, with the "discharge momentarily and then let "recover" method", where the "recovery voltage" is measured after 15 minutes, which again is probably one or two orders of magnitudes away from the actual point where the "recovery" reaches its maximum).
In other words, it's quite likely, that the few thousands of MOhm*uF figures given by the manufacturers for the ceramics, may be one or two orders of magnitude off, and that the real "leakage time constant" may be around or above our target 36ks.
As it was only on grounds of these leakage figures we dismissed the ceramics as candidate for a night-long storage, we now can consider them as candidates again. Thus, we can consider the simplest arrangement, the capacitor being charged from the primary source through a low-leakage diode to some 3V and been connected to the RTC directly.
Our estimated capacitance need for the super-low consumption RTC is a few hundreds of uF. 100uF X5R MLCC rated for 4V are readily available in 1210 size for around a dollar apiece and maybe half a dollar at moderate bulk. This is a very reasonable price to pay for a gadget to last a century.
There is hope.