What Dreaming Actually Does
The name comes from how the process runs: in the background, between your conversations, the way sleep consolidates memory in biological systems.
OpenAI describes dreaming as a method for ChatGPT to automatically curate memories by referencing chat history. It synthesizes facts across many conversations into a memory state that persists between sessions. When you come back, the model has a compressed representation of what it learned about you, without you having to paste context or remind it of past threads.
The improvements are genuine. Before Dreaming, memory was limited by what the model could explicitly hold and often redundant, with multiple near-identical entries from different conversations. The synthesized approach consolidates these, updates facts when they become stale, and trims entries that no longer apply. A trip you were planning in March is recognized as a past trip by July, rather than staying forever in the future tense.
The compute argument is also genuine. Running memory for hundreds of millions of users requires inference at scale, and summarizing rather than storing verbatim reduces both what the model has to hold and what it has to process per request. OpenAI reports a roughly fivefold reduction in compute required to serve memory to free users, which is the kind of number that determines whether a feature ships broadly or stays gated to paid tiers.
What You Can and Cannot Control
The settings page still exists. You can view the memory summary, delete individual entries, and turn memory off entirely. These controls are real and they work. What changed is the direction of flow.
Before Dreaming, the default was forgetting: nothing was remembered unless something triggered a save, and you could explicitly ask it to remember or forget. After Dreaming, the default is remembering, with curation: the system runs continuously, decides what matters, and synthesizes a persistent state. You can delete what you notice, but you are reacting to a process you did not initiate.
User feedback on the OpenAI forums reflects the tension. One thread titled 'Dreaming V3 Sacrifices Real Personalization for Compute Savings' argues that the rollout removed user control over what gets saved and that a background process now synthesizes and curates memories without user input. The same thread notes that users can switch back to saved memories in settings, but the framing matters: the new default is automatic, and the old behavior is now an opt-out.
This is not a complaint that the feature is bad. The recall improvements are real, and for most people the automatic approach probably works better than explicit memory management, which few users bothered with. The point is narrower: if you think of what an AI knows about you as something you should control, Dreaming moves the locus of control from you to the system, and calls it an upgrade.
Why This Matters for Your Actual Notes
The appeal of AI memory is that you do not have to remember to remember. The cost is that you also cannot decide what gets kept.
Consider what happens when the system synthesizes incorrectly. A fact you mentioned in passing becomes a stored belief. Two separate projects get merged into one because they shared a keyword. An offhand complaint becomes part of your profile. You will not know this happened unless you read the memory summary, and the summary is itself a compression, so even reviewing it does not show you the raw material it was built from.
The deeper issue is not accuracy. It is that synthesized memory is not your notes. It is someone else's summary of your notes, stored in someone else's system, updated on someone else's schedule. You cannot grep it. You cannot version it. You cannot export it to another tool and have it mean the same thing. It is useful exactly as long as you use ChatGPT and exactly as long as OpenAI offers this feature, and those are not conditions you control.
A second brain is supposed to extend your own. The whole point is that it holds what you chose to keep, in the form you chose to keep it, available when you need it. Dreaming offers a different proposition: a second brain that curates itself, that you can consult but not govern, and that lives inside a product rather than under your roof. For some people that is a better deal. For anyone who cares about owning what they know, it is a warning sign wearing the clothes of a feature.
| Your notes | AI-curated memory | |
|---|---|---|
| Who decides what to keep | You | The algorithm |
| What you see | The source material | A summary |
| When it updates | When you write | In the background |
| Exportable | Yes, if the format is open | Not in any useful form |
| Survives the product | Yes | No |
| Correctable | Edit the file | Delete and hope |
The Memory You Cannot Export
OpenAI does offer a data export. You can download your chat history as a set of files. What you cannot export is the synthesized memory state itself, the thing Dreaming built, in a form that another tool could use. The memory exists inside ChatGPT, as a feature of that product, and it does not come with you.
Anthropic's Claude offers a memory import feature that accepts exports from ChatGPT, Gemini, and Grok. But what it imports is the raw chat history, not the curated memory, because the curated memory is not available as a file. The synthesis happens again, from scratch, inside Claude's system. Your profile in Claude will not match your profile in ChatGPT, even if both were built from the same conversations, because the curation is proprietary.
This is the portability problem restated for the memory era. In the old world, vendor lock-in meant your documents were stuck in a format only one app could read. In the new world, vendor lock-in means your AI-derived identity is stuck in a system only one vendor can interpret. The documents are exportable. The understanding is not.
The Pattern to Recognize
Dreaming is one example of a broader shift. AI assistants are moving from tools you direct to systems that anticipate. Proactive suggestions, background processing, automatic personalization: the common thread is that the system does more and you decide less. For many tasks this is an improvement. Remembering to tell your assistant what you already told it is friction nobody will miss.
But knowledge management is not like other tasks. The value of a second brain is not that it remembers for you. It is that it holds what you chose to keep, retrievable later, trustworthy because you put it there. Automatic curation breaks that loop. You might find what you need, or you might find what the algorithm decided was important, and the two will often but not always overlap.
The practical response is separation. Use AI memory for what it is good at: continuity across conversations, personalization within a product, convenience when convenience is the goal. Keep your actual notes somewhere else, in a form you control, exportable in formats that outlive any one tool. Let the AI read your notes when you want its help. Do not let your notes be the AI's memory, because then they are not yours.
Mindly keeps the notes on your Mac and lets AI help with them, rather than turning your notes into AI training data or background curation material. A library you control →
What to Do Now
- Review your ChatGPT memory summary. Go to Settings, then Personalization, then Memory. Read what it synthesized about you. Delete anything that is wrong or that you would rather not have stored.
- Understand the default. Dreaming is on unless you turn it off. If you want explicit control over what gets saved, you can switch to saved memories mode in the same settings panel, but automatic curation is now the baseline.
- Do not rely on AI memory as your archive. Treat it as a convenience layer, not a source of truth. The record you might need in two years should live in your own files, not in a feature that might change or disappear.
- Export your chat history periodically. OpenAI lets you request a data download. Do it on a schedule and keep the archive somewhere you control, because the synthesized memory is not exportable.
- Keep your actual second brain separate. Use AI to help you capture, tag, and retrieve. Do not let AI be the thing you are retrieving from, because then you are retrieving from someone else's summary of what you said.