1/
A memory UX request from someone who’s been using ChatGPT since the first months after launch in late 2022.
I think part of my frustration comes from a mental model I built during the GPT-3.5 / early GPT-4 era, when 4K and 8K context windows were normal.
Back then, New Chat meant a true clean slate: wipe the context, return to the default model state, and start over.
Sometimes you actually needed to do this, because a long or messy context could noticeably degrade response quality.
After using ChatGPT that way for years, “New Chat = clean slate” became muscle memory for me.
2/
When the original Saved Memory system arrived, that mental model changed a little — but it still felt very transparent and controllable.
When ChatGPT read or wrote memory, the UI made it obvious with indicators like “Recalled memory” or memory update notifications.
Saved Memories were also discrete objects that I could inspect and delete directly.
So even if ChatGPT suddenly knew something from another conversation, it usually didn’t bother me. I knew why it knew, I knew memory had been involved, and I could go into Saved Memories and remove the relevant item if I wanted to.
3/
The new memory system is genuinely impressive, and most of the time I really like it.
Automatically synthesizing my history and surfacing relevant context without me having to manage memory manually is a huge improvement. Most of the time, the fact that it feels invisible is actually a feature.
But that seamlessness also created a new problem for me.
I once opened a brand-new chat specifically because I wanted to retry something independently after getting a strange result in the previous conversation.
ChatGPT immediately responded with something along the lines of:
“This time it (the value) isn’t X like last time — it’s Y.”
Technically, that was useful memory.
But subjectively, it was a surprisingly uncanny moment. I had deliberately opened a new window because I expected an independent run, yet the model was clearly carrying information across the conversation boundary.
For the first time, “New Chat” no longer meant what I had expected it to mean since 2022.
4/
Temporary Chat gets very close to solving this.
I especially like that it now has a Personalization toggle, so I can choose whether a temporary conversation should use my existing personalization/memory.
That part makes sense.
The problem is that memory isolation is still coupled to the lifetime of the conversation.
A Temporary Chat is not a normal persistent conversation sitting safely in my chat history. If I refresh at the wrong time, close the app, lose the session, or simply want to come back to it days later, it doesn’t behave like a normal saved chat.
But these are two completely different questions:
“Should this conversation use my memory?”
and
“Should this conversation still exist tomorrow?”
I don’t think one should imply the other.
5/
What I’d really love is a persistent memory-isolated chat — essentially a persistent version of a non-personalized Temporary Chat.
Something like:
- Save to chat history: ON
- Read from global memory / past-chat personalization: OFF
- Write back into long-term memory: OFF
The conversation itself would behave like any normal chat:
it survives refreshes, app restarts, device changes, and I can come back to it days or weeks later.
But its relationship with global memory would be completely isolated in both directions.
6/
The “don’t write to memory” part is especially important to me.
This isn’t only about sensitive information.
Sometimes I want to talk about something that is:
- too trivial to belong in long-term memory
- a temporary preference
- hypothetical or deliberately unrepresentative of me
- part of an A/B test or model evaluation
- an experiment where I specifically want the default model with no personalization
- or simply something I don’t want affecting future conversations
I might still want to spend hours on that conversation and return to it tomorrow.
I just don’t want it becoming part of ChatGPT’s long-term understanding of me.
7/
Right now, if I want this kind of persistent “clean room,” the workaround is awkward:
I have to go into Settings, disable my global Memory / Personalization settings, start the experiment, and then remember to turn everything back on afterward.
That feels unnecessarily global for something that is really a per-conversation choice.
And if I forget to restore the settings, I’ve changed the behavior of every other chat just because I wanted one isolated conversation.
A per-chat option would remove that friction entirely.
8/
To me, there are really three independent properties:
Conversation persistence ≠ Memory read ≠ Memory write
Temporary Chat already proves that ChatGPT can isolate a conversation from personalization.
What’s missing is the ability to keep that isolation without forcing the conversation itself to be temporary.