#gpt-3.5-turbo forgets quickly

4 messages · Page 1 of 1 (latest)

heavy bison
#

I'm using GPT3.5 for my bachelor thesis but after testing it a while it seems like it keeps forgetting messages.

I'm sending the history of my last 10 messages and the starting message (for instructions) but if the history reaches a certain length it keeps forgetting things in the instruction. I can also reproduce this behaviour in the playground.

Is this normal and is there a way to avoid forgetting certain messages or is the only way to trim the history that is send to the API?

dawn widget
#

Are you sending context? I send like the last 10 prompts and responses?

heavy bison
#

Yes i'm sending the history of the last 10 responses (with the first initial 'your role is to be Mark') but as soon as the history gets too much text i cant ask for the name or any other stuff that was mentiont in the initial message.

Even the playground has this problem. As you can see in the image i can ask for the name and what i need to do. But in user message 6 it forgot its role and gives the default answer 'As an AI language model ....'

#

I can avoid this if i keep my history short. Like 4 user inputs with responses seems to be ok, but anything above that makes problems.