I can not understand how this hasn’t been capitalized on more yet! Companies with the data MUST be training on it as RLHF, right?
In general, the “chat history mining” sector seems to be way under-developed to me so far.
All the damn time I am annoyed I have to re-tell my LLM a piece of info I have already told it a few weeks ago - finding the chat takes too long and the full history may not be relevant, but the fact wasn’t interesting enough to manually add to memory.
With the right memory system, the LLM could just know!
Does anyone have other interesting examples of this principle being applied?
we did the same thing, but we noticed most users did not maintain their own system memory themself properly. so we had to build a solution that auto-manages memory and preferances.
> user corrections are the highest-signal data
I can not understand how this hasn’t been capitalized on more yet! Companies with the data MUST be training on it as RLHF, right?
In general, the “chat history mining” sector seems to be way under-developed to me so far.
All the damn time I am annoyed I have to re-tell my LLM a piece of info I have already told it a few weeks ago - finding the chat takes too long and the full history may not be relevant, but the fact wasn’t interesting enough to manually add to memory.
With the right memory system, the LLM could just know!
Does anyone have other interesting examples of this principle being applied?
Note that you posted an “Ask HN” (with it linking to this page) - do you want to show us something instead?
for now we just use normal system memory the user can maintain himself - not the best solution but better than nothing.
we did the same thing, but we noticed most users did not maintain their own system memory themself properly. so we had to build a solution that auto-manages memory and preferances.