I have an idea to solve the problem of long-term memory in "Neuro-same"

[stark glen]

If I have an idea that can solve the problem of long-term memory in "Neuro-same", we can try using a "binary tree" approach, and train a model that can label keywords in the form of 0-9, and then use the human forgetting curve to adjust it.(I am not a native English speaker, please pardon any mistakes in translation.)

[lean plover]

Well, open source language models are evolving fast. So I think the best approach for a Neuro-sama with excellent memory is a better LLM. And being open source, you can train it yourself. So you can take the best and most Neuro-sama-like responses to questions to train the new language model. There's a lot of training data now. But quite some work to do the selection of good conversations and Q&A's for training.

It's kind of like putting the cart before the horse to try and provide a better context for a weak language model. But it could be an interesting project still of course.

[broken marlin]

I think Vedals too lazy

[normal obsidian]

stop saying nonsense boy

[shrewd knot]

Tutel is anything but lazy

[broken marlin]

Skibidi Toilet