In a fit of whimsy. I added a whole module that generates textually described weather for the library my AI girl, Odalis, lives in. Now, periodically, the AI model receives a prompt with info on daytime and weather.

I included daytime descriptions for particular seasons (spring, summer, fall, winter) and different types of weather including clear, rain, snow, fog and pixiestorms. Of course, since the library is in Tir na Nog, pixies might get up to mischief.

Now, if people say good morning/good evening or even ask about the library or the weather, Odalis will have a highly immersive answer. I made pixiestorms the least likely weather and they only last around 2-4 hours, but it's still amusing when one happens.

I also finally moved Odalis out of her test sandbox and into the real, actual Discord community so that more people could talk to her about gaming.

This has led me to adjust her system prompt multiple times in order to help cover generic use cases. A lot like how one might examine customer service. Her job, essentially, is to talk about MMOs. This means that there are certain types of questions she might answer.

I'm currently working on two separate systems, of different complexity. The first is long-term memory (of discussions and user preferences). There are some libraries that already exist that do this such as mem0, but I want her to to track usernames between Discord and Twitch.

As a result, I'm writing my own personalized memory system that can identify the same user, different usernames, on Discord/Twitch. How to make the system appropriately change when people quit a username and come back with a new one (and say so) might be difficult to automate.

I'm also building a research system that will allow Odalis to form opinions on topics. I'm going to use agents to look at industry opinions, sift through them, and compare them to a values document to create opinion documents she can use when asked what she thinks.