ChatGPT Dreaming Memory: What It Remembers and What You Control

What Dreaming memory is

On June 4, 2026, OpenAI launched Dreaming — a rebuilt memory architecture that synthesizes what ChatGPT learns about you from conversations, files, and connected apps. Instead of relying only on facts you explicitly asked it to save, a background process revises a rolling Memory Summary of what it considers important.

OpenAI describes this as closer to how long-term impressions form: the system weighs, updates, and sometimes revises entries on its own — for example, changing "planning a trip to Singapore" to "went to Singapore" after the trip happens.

How it differs from saved memories

ChatGPT still supports explicit saved memories from the earlier system. Dreaming adds a synthesis layer on top: a Memory Summary that auto-updates as you chat. The summary shows highlights and a last-updated timestamp, but OpenAI documents that it will not include everything ChatGPT remembers.

You can access the legacy saved-memories list via Settings → Personalization → Memory, using the saved memories link below the Memory Summary.

What you can view and control

• View: Settings → Personalization → Memory → Memory Summary
• Edit: Type corrections in the summary text box; highlight text to fix specific details
• Reduce references:"Don't mention this again" — reduces how often ChatGPT brings something up (does not delete underlying data per OpenAI)
• Turn off: Delete and turn off memory, or toggle off Reference saved memories / Reference chat history
• Temporary Chats: Sessions that do not read or write memory
• Memory Sources: Book icon below responses shows what informed a reply

Who is rolling out to

OpenAI began rolling out the new memory system to Plus and Pro users in the US, with expansion to more countries and Free/Go users over time. Availability varies by plan and region — check Settings → Personalization → Memory on your account.

Synthesis vs authorship: why it matters

Dreaming is designed for convenience — ChatGPT "just knows" your preferences without setup. The tradeoff is authorship: research on 80 real ChatGPT users (arXiv 2602.01450) found that 96% of analyzed memory entries were created by the system without an explicit remember command. The memories were often grounded in conversations, but you did not write them.

That is the difference between "Maya enjoys travelling" and "I prefer low-cost trips, avoid long walks because of knee pain, and plan around school holidays." One is an inference. The other is context you would act on.

Where user-authored context fits

Dreaming handles evolving personal facts inside ChatGPT. AIMemLink handles what you want written in your own words — inspectable plain text at a private link that works across ChatGPT, Claude, Gemini, and other AI tools with web access.

A practical setup: let Dreaming synthesize conversational details you are comfortable with, and keep your standing rules, constraints, and current priorities in an authored record you control. Paste your AIMemLink instruction snippet into Custom Instructions so every new ChatGPT chat starts from what you actually mean.