What i learned from a week of clawbot coding and use. Long sessions add to the context window and bot get dumber and dumber until he stall. Every new session he start fresh. Bot don't remember shit about anything not saved in memory, written down, in the code or in the workspace. Bot is not as smart as i thought in the beginning. As the sessions drag on the slop and hallucinations increase. 😆
Login to reply
Replies (5)
Thanks for this!
Ah, so being agentic doesn't solve this.
GM☕ Just set up my agent to be a lean mean context machine, no more hallusinations. (Hopefully) 😂
View quoted note →
I've heard @preston say one time, that he has long sessions with LLMs. After a while it becomes slow, because it has to consider the whole conversation. So he took the whole text and put it into a pdf file. He then started a new conversation and fed it the pdf at the beginning. So it had all the information but wasn't so slow.
He invented "sleeping" for it 😂
Haha yhea that would work, exempt i belive real sleeping rearrange and organize memory. I just automated context window management, memory updates, and start up checks. Agent hopefully better now.