OpenAI just gave you a window into what the model remembers about you, then immediately admitted the window has curtains you can't see behind.
The Summary
- GPT-5.5 Instant replaces GPT-5.3 Instant as ChatGPT's default model, promising better accuracy and fewer hallucinations across all free and paid tiers
- New "memory sources" feature shows which context shaped responses, but OpenAI admits models "may not show every factor that shaped an answer"
- Users can now tap a sources button to see which files or past chats informed the model's response, and delete outdated context
- The partial transparency creates a second, incomplete observability layer that could conflict with enterprise audit systems and agent logs
The Signal
OpenAI's new memory sources capability sounds like transparency until you read the fine print. When ChatGPT personalizes a response, users can tap a sources button to see which saved memories or past conversations the model referenced. You can delete or correct outdated information. Full control over your digital footprint, right?
Not quite. OpenAI explicitly states the models "may not show every factor that shaped an answer" and promises to make the feature more comprehensive over time. Translation: the model knows more than it's telling you. This isn't full observability. It's curated observability.
"When a response is personalized, you can see what context was used... and delete or correct it if something is outdated or no longer relevant."
For enterprises, this creates a problem. You now have two memory systems running in parallel:
- The visible layer: what ChatGPT shows you in the sources button
- The invisible layer: whatever else shaped the response that the model won't or can't reveal
- Your existing audit logs and agent monitoring tools, which may capture different context entirely
The GPT-5.5 Instant model itself represents a meaningful step up from GPT-5.3 Instant. Reduced hallucinations, better accuracy, smarter reasoning. These improvements matter for production deployments where reliability is non-negotiable. Making this level of capability free for everyone accelerates the baseline expectation for what AI should do.
But the memory sources feature reveals something more interesting about where we are in the agent economy: we're past the "does it work" phase and deep into the "can we trust what it's doing" phase. Users don't just want better answers. They want to know why the model said what it said, what it remembers about them, and how to correct the record when it's wrong.
Key tensions in partial memory observability:
- Users get visibility into some context, not all context
- Enterprises need full audit trails for compliance and debugging
- Agent systems may make decisions based on factors humans can't inspect
OpenAI's admission that this is incomplete suggests they know full observability is technically hard or strategically undesirable. Maybe showing every weight update and context window priority would expose too much about how the model works. Maybe it would overwhelm users with noise. Or maybe they haven't solved it yet.
The Implication
If you're building agents on ChatGPT, don't assume the memory sources button shows you everything. Treat it as supplementary context, not ground truth. Keep your own logs of what you feed the model and what it returns. When personalization goes sideways, you'll need your own paper trail.
For individual users, the sources button is still useful. You can finally see when ChatGPT is pulling from that conversation you had three months ago about your cat's diet and correct it when the cat is now on a different food. That's real. Just know there's more happening under the hood than what you see. The model is making calls you can't audit. Plan accordingly.