Why don't Custom GPTs remember user preferences? Isn't that the point?

ChatGPT
1

I’m using ChatGPT Plus to create a Custom GPT, but the GPT is telling me that it can’t remember user preferences between sessions. Why else would I want to create a Custom GPT then?

For instance, I am using one to generate music playlists and I want the output format to be ARTIST - TRACK (no “” quotes) , but why can’t it remember this? If I start a new session, it defaults to ARTIST - “TRACK”. Or even sometimes ARTIST - “TRACK” (YEAR)

I also want it to never suggest certain bands such as Arcade Fire and Band of Horses, but it can’t remember this either.

If Custom GPTs can’t remember simple preferences like this, then what’s the point? How is it any different than regular ChatGPT?

1 Like
2

Sessions are identical to conversations.

You would need to implement the restrictions and formats you described in the GPT builder so that it’s persistent across all instances of the custom GPT.

If you think of this like an app, consider what preferences you want the user to be able to adjust in very precise and specific detail. If the user can adjust whatever they want, then yes, it’s no different than standard GPT. However, I will tell you the power in custom GPTs lies in actions. Building actions are how you can make the model clearly distinct from its source material.

3

I guess where I’m confused is that I put some “Instructions” into the custom GPT builder for music playlists, but it often outputs songs that don’t conform to the instructions that I gave it, such as to focus primarily on deep tracks, b-sides, and discovery instead of popular tracks. These instructions are embedded into the GPT via the Builder, I am not instructing it during a conversation.

So after it gives me a playlist that doesn’t fit the instructions (e.g. has too many popular tracks) I will say “Are you sure that this playlist meets my specifications?” and it will say:

Upon reviewing your request and the playlist I provided, I realize it may not fully align with the initial instructions for creating a playlist that offers a diverse and unexpected musical experience. The playlist I crafted leans heavily on well-known tracks and popular artists from a range of alternative and indie genres, but it may lack the deep cuts and less mainstream selections that would truly surprise and engage a listener seeking diversity and eclecticism.

Let’s try again, focusing more on deep cuts, lesser-known tracks, and a broader variety of genres, while still avoiding the specified artists

THEN it outputs a playlist which mostly conforms to those instructions, but it has ANOTHER problem: it includes artists which I explicitly instructed it never to include in the playlists (when creating the GPT in the Builder, and also in the Instructions field for it).

So then I say, “Doesn’t this playlist include artists which I told you not to recommend?” and it responds:

You’re absolutely correct, and I apologize for the oversight. Let’s create a new 100-song playlist, this time carefully avoiding the artists you specified (Arcade Fire, The Decemberists, Band of Horses, Neutral Milk Hotel, Grimes, The Knife, Fever Ray) and focusing on providing a diverse and eclectic mix of genres and eras:

Why is it forgetting its instructions, yet remembers the instructions when I ask why it’s giving me output which doesn’t conform to the instructions?

This seems like a massive oversight…

1 Like
4

I’ve encountered this same issue before with custom instructions. It, sometimes, does what it wants.

With GPT Builder, I’ve found it better than custom instructions but around 20% of the time it deviates significantly from an instruction, even when explicit.

5

I have the same issue. I’ve set up my customized GPT, and instructed the GPT builder to perform a specific task. Initially I’m happy with the results however as soon as I start a new conversation it deviates greatly from my instructions. Is there a way to keep it consistent?