#GPT-4o loses key instructions in long conversations, what's the best way to remind it?

1 messages · Page 1 of 1 (latest)

primal narwhal
#

Hello everyone,

I've been working on a custom assistant using the OpenAI playground for months, and I'm consistently running into a major issue. My assistant works great for the first couple of messages, but as the conversation gets longer, it starts to forget key instructions from my initial prompt.

I'm using GPT-4o with a temperature of 0.10, so the responses are very consistent, but the core problem of it losing focus persists.

I've read some articles and posts that suggest techniques like "restate, remind, and frame" within the conversation, or even adding reminders to the "run" instructions. The latter seems like a poor practice, but I'm at a point where I'll try anything to make this work.

Has anyone else faced this exact problem with the model losing its "north" in longer chats? What was your solution? Is there a best practice for managing memory or reminding the model of its core instructions without completely disrupting the flow? Any help or examples would be greatly appreciated.

Thanks in advance!

vapid falcon
primal narwhal
#

in the system instructions, i read that i could add the whole prompt in the run instructions but is that a good practice? if i add it in the run instructions what do i put in the system instructions?

vapid falcon
#

it should work fine with system instructions. did you try the same instructions with responses api and few messages? if this only happened at one point then check if the model got updated at the same time if you use the main alias