/vt/ - Virtual Youtubers

>>50393848
From what I understand, she basically has short time memory and global memory.
Short time memory is whatever she is currently talking about, recent responces and details etc. I also believe that when talking to chat, whenever she goes on a rant her short time memory partially resets.
Global memory is language model database. All data fed to it gets processed and added to the main pile, to later be used as a basis to generate responces. In other words, it takes lots and lots of text to try and generate the most common responce(with some randomness apparently). It means that anything added to Neuro's memory, whether it is random script or her own collab, she will not perceive as her own memory and will just use as an example. So theoretically, if Vedal adds full script of Neuro's streams and collabs to her memory, she will "learn" from herself and will be more likely to give out same lines as before. The more she repeats herself, the more she will in the future. This is probably also why she mentions Vedal so much, he appears in her memory way more than any other name. And I'm pretty sure her "KEKW means laughing" and "Flushed means blushing" is Vedal's attempt to forcefeed her context of emotes, but she just took it as raw value, simply repeating the same phase over and over, though recently she started actually trying to use them in context.
It also means that there's a lot of data there that is not usually accessed, but is still there and can potentially be brought up with a correct request. Like, remember the 10 tin cans 1 stream moment? Back then Neuro's memory was still pretty raw and she could bring more random shit without context as it had more or less same weight as other random shit. Now it is more refined and she gives more consistent and coherent speeches because she has more examples of what is more commonly used.