Letta's reasoning feels strange to me

[meir23]

Hello everyone

Is the reasoning step done using the function: _get_ai_reply? And is it the product of a language model such as: gpt or claude?

I ask this because I feel that raising the ability to understand based on context is very, very bad in relation to the tests I do for it.

For example: if I make some request >> the reasoning limits it or does not approve it (probably because of the system prompt) >> and then I give a continuation command (like in normal speech between people where there is a dialogue and what I say is in response to what you said) then the system responds poorly. That is, it feels to me in many cases much stupider than using gpt and/or claude.

Does anyone else share these insights?

If I may, I'll give a follow-up question... Can I simply build a tool that will do the reasoning and have a rule that will always run first + after every other tool?

And can it eliminate the reasoning itself (or as it was once called, internal thoughts)?

In my opinion, it feels like the reasoning is really an integral part of the system

[mruderman]

What model are you using? There are a number of models that have come out for Letta using reasoning oriented approaches (see https://huggingface.co/mradermacher/Letta-o1-i1-GGUF for an example).