this aspect of honcho is primarily focused with the maintenance of the user representation. people are complex, ever-changing beings… the user representation needs to remain just as plastic.
Post Raise Summer ‘24
Claude suggested iteratively updating the user vector with a weighted average of the successful thoughts. I’m sure there’s a simple time decay variable to add to this weighting, but it wasn’t explicitly discussed.
Historical Recap
Up and to this point (2024-07-15), maintaining the user representation has been somewhat overlooked. In the violation of expectation approach, we would check the existing user representation to see if facts we had derived already existed and not write them if they did. However, there wasn’t really any mechanism to do a wholesale update to the representation. Nor did we have any sort of time-decay or weighting going on. There is a lot to improve… basically no feedback loop.