Before Ai was a thing, I created a little spreadsheet app to simulate a type of stochastic memory. I created a "database" consisting of places and their capitals. The database began with no entries. I would input a place and ask the app for its capital. The app would select a random number from 0 to one, multiply it by the number of entries in the database, round to the nearest integer, and see if the entry in the database with the corresponding number matched my query.

On the first try, I would enter France, the app would generate 0 (because there were no entries in the database), and the test would generate an error (because there was no such thing as a zeroth entry). The app would try again for a certain number of tries (the "budget") and then give up. At that point, I would enter France and Paris side by side in the database. Then I would enter United Kingdom and see if the app could find its capital by randomly looking at database entries (which, at this point consisted solely of "France"). It could not find the UK within the budget (because it wasn't there), so the UK, London pair was added to the list,too.

Eventually, I'd have a lot of pairs. Then, if I asked for France, the system would randomly query the database a number of times. If the budget was adequate, one of the trials would land on France, and the app would return "Paris." But if, after exhausting the budget, the app failed to hit on France, another "France, Paris" pair would be added to the database, doubling the probability that I would find it next time. That way, the things I needed to know most often would have the most entries and be easiest to retrieve within the search budget.

I never pursued the matter far enough to optimize the search budget, but I could see ways in which it could be made a function of the number or entries in the database, the redundance of common queries, or even the importance of the search. I think I was on to something...