This is probably table stakes for most of you, but I thought I’d pass this tip along as it may be helpful to someone.
I instructed ChatGPT to place in its memory to search the web for the latest support pages prior to answering any technical questions. After giving this instruction, I have found that the answers I receive back when I have a technical question are more accurate. The key is to tell ChatGPT to always search the web for the latest support information prior to answering a technical query. Once ChatGPT has this in memory, you don’t need to repeat the instruction. However, for my first three queries after giving the initial instruction, I prompted ChatGPT, “Remember to always search the most current support pages prior to answering technical questions.”
Among the popular LLMs, only Chat GPT has persistent memory across chats. Claude allows you to setup Projects and add reference material that can be used throughout chats in that project.
And Perplexity does the same thing you describe with Claude ie. ‘spaces’ where you can provide reference materials and a custom prompt (but not ‘memory’ afaict).
modify ChatGPT’s system prompt (which is under Personalization → Customize ChatGPT) and
modify ChatGPT’s memory (which is under Personalization → Memory)
The difference is that the system prompt will always be transmitted to the model while (I believe) parts of the memory are used included based on some RAG method, i.e. only memory snippets supposedly relevant for your current prompt will we sent to the model.
I think it is better to add anything you want ChatGPT to always remember to the system prompt.
Does anybody know more about how exactly memory retrieval works in ChatGPT?
Hmm. I’d had an ongoing issue with ChatGPT reading sheet music. I have a long and ever changing list of church music in PDF form from which I need to rip the lyrics only, and in correct verse order. It’s a real pain for me to do this manually. And as folks familiar with church music know, there are often many versions of the same song, but just slightly altered enough to make it “unique.” This usually comes in the form of rearranged verses and " you and yours" exchanged with “thee and thou,” etc. Super frustrating. Anyway, I found that after uploading the sheet music, with explicit instructions to read the upload only, ChatGPT would “cheat” and find the song online somewhere and retrieve the lyrics that way. In almost every incidence, the retrieved lyrics were from a slightly different version than the one I uploaded. I kept pointing this out to ChatGPT and rephrasing the request, but it always came back the same way - incorrect. Oddly, it would always apologize and admit how it got the lyrics; but it would repeat the same method regardless.
I ended up getting Grok 3 to do a much better job at what I wanted, and although it gave me a bit of grief too, it still remembered the detailed instructions the next time around.
