I'd be surprised to hear that lots of people don't do just this. As soon as the memory features came out I gave them a quick try and quickly turned them off.

I don't want to be held accountable to all of my previous ideas. I want each conversation to start fresh with the context that I provide. If I am exploring some library in one language stack and then I later want to look into something completely different, I don't want the conversation polluted by what it thinks I want based on the previous discussion.

I suppose for those who use it as a companion the memory is a core element. But when used as a tool it gives a significantly worse experience IME.

Memories inside projects are a boon: If you can assume that all your questions are related, and you instruct the engine what memories to store and how to store them (short single statements), you can get a powerful tool out of it. I use it for habit tracking. The system prompt is taking the style of Atomic Habits. It will remember exactly what habits I'm trying to add, what my preferences are for apps, when I sleep, what other goals I have unrelated to my current query.

Concrete memories in a generic context suck:

Every time I ask Claude Kubernetes question, it relates it to that one time I was deploying Kubernetes on VMs. It's not helpful!

Every time I ask Claude a question related to projects, it starts mentioning where I work, as if that helps.

Asking it to forget these concrete memories sometimes results in remembering to not mention them.

Thoroughly useless.

I started out disabling it because it's creepy. I tried it out, and it just doesn't work well.

I rather have memory and ideas in Markdown and pick what to include in the context rather than relying on platform specific memory features

Leaving memory enabled risks introducing bias and baggage from past chats and experiences that are better off being left behind. It however then becomes my responsibility to tell ChatGPT everything it needs to know in every chat.

Note that there are other sources of bias besides memory. For example, for each Custom GPT, an OpenAI staff Karens looks to type up and insert custom instructions that apply with a higher priority, and often are detrimental to the output, thereby sabotaging the Custom GPT. It would be better off if they didn't do any of it. I have had multiple Custom GPTs that initially worked flawlessly get sabotaged by OpenAI in this way.