ChatGPT Dreaming V3: How OpenAI's New Memory Architecture Works and What It Means for Free Users
OpenAI's Dreaming V3 replaces ChatGPT's saved-memories list with background synthesis, doubling factual recall and bringing smart memory to free users for the first time.
ChatGPT’s memory just got a significant overhaul. OpenAI has launched Dreaming V3, a new memory architecture that replaces the manually curated saved-memories list with a background synthesis process that reads across your conversation history and keeps itself up to date without you needing to do anything. And for the first time, it’s coming to free users too.
Here’s what’s actually changed, how it works, and what it means in practice.
The problem with saved memories
When ChatGPT memory first launched in April 2024, it worked roughly like a notepad. During a conversation, if you said something worth remembering, or explicitly asked ChatGPT to remember it, a short text entry got saved to a list. Future conversations could reference that list.
The obvious limitation: it only captured what you told it to capture. Context that came up naturally in passing often went unrecorded. And entries went stale. If you told ChatGPT you were planning a trip to Singapore in July, that fact would sit in the list indefinitely, still phrased as future tense long after you’d come home.
In April 2025, OpenAI introduced the first version of dreaming, which let the system pull context from chat history beyond the saved list. Dreaming V3 takes that much further and, according to OpenAI, replaces saved memories as the primary foundation of the memory system entirely.
How the background synthesis process works
The core mechanism is an asynchronous background process that runs outside your active conversations. It reads across your chat history, synthesises what it learns about you, and continuously updates a memory state that gets injected into the system prompt at the start of each new conversation.
This means the system can pick up context that you never explicitly flagged. If you mentioned in passing that you have a gluten intolerance, or that you’re learning Spanish, or that you started a new job, that information can find its way into your memory profile without you having to say “remember this.”
Crucially, it also updates over time. OpenAI’s own example: a memory reading “you’re going to Singapore in July” rewrites itself to “you went to Singapore in July 2026” once the trip is over, and recommendations shift back to your home context automatically. With the old saved-memories approach, that kind of staleness correction required manual intervention.
The numbers: factual recall from 41.5% to 82.8%
OpenAI published internal benchmark figures comparing its memory systems across three years. On factual recall, task success went from 41.5% with 2024 saved memories, to 67.9% with the 2025 system, to 82.8% with Dreaming V3.
Preference adherence followed a similar trajectory: 31.4% in 2024, 55.3% in 2025, 71.3% now. The most striking improvement is in time-sensitive accuracy, which tests whether the system correctly updates outdated context. That started at just 9.4% in 2024, reached 52.2% in 2025, and now sits at 75.1%.
These are OpenAI’s own internal evaluations, and the methodology and datasets haven’t been published, so treat them as directional rather than independently verified. But even with that caveat, the trajectory is notable, and the time-sensitive accuracy improvement in particular addresses one of the most practical complaints about how AI memory has worked until now.
The new memory summary page
Rather than a flat list of saved text snippets, you now get a memory summary page that shows you what the system has synthesised about you. From there you can add information, correct things that are wrong, give ChatGPT instructions about topics to focus on or avoid, and delete individual entries at any time. If you want to dig into a specific area, you can chat directly with the model about it.
One thing worth knowing: OpenAI acknowledges the summary does not necessarily show everything the system may have inferred from past conversations. The memory state lives in a separate data layer from your conversation history and is injected at inference time. Selecting “Don’t mention this again” on a summary entry reduces future references but does not delete the underlying data. You need to explicitly delete an entry to remove it.
If you prefer the old saved-memories behaviour, you can switch back in settings. Temporary chats do not use or update memory, and you can disable memory entirely.
Free users get real memory for the first time
This is the part of the announcement that matters most to the largest number of people. Until now, dreaming-based memory has been exclusive to Plus and Pro subscribers. Free users had access only to the older, more limited saved-memories system.
What changed is compute efficiency. OpenAI says it reduced the processing cost of running dreaming for free-tier users by approximately 5x through architectural optimisations in how the background synthesis is batched and served. The specific technical method hasn’t been published, but the practical result is that a process previously too expensive to run at free-tier scale is now viable across the full ChatGPT user base.
For Plus and Pro users, the same efficiency gain translates into roughly 2x more memory capacity.
The rollout is live now for Plus and Pro users in the US, with additional countries and free-tier users following over the coming weeks.
What this means for you
If you use ChatGPT Plus or Pro, the most immediate change is that ChatGPT should feel more consistently aware of your context across sessions, with less drift over time. You no longer need to manage a list or prompt the system to record things. The memory summary page gives you a cleaner way to see what the system knows and correct it when needed.
If you’re on the free tier, this is the first time you’ll have access to memory that actually learns from your conversations in the background rather than just storing what you explicitly told it to save. It’s arriving over the next few weeks.
The open question is auditability. The shift from an explicit list to an inferred memory state makes the system considerably more capable, but also less transparent about exactly what it knows and why. The memory summary page helps, but the acknowledgement that it doesn’t surface everything is worth keeping in mind. If that matters to you, the option to disable memory entirely remains available in settings.
OpenAI describes Dreaming V3 as “a shared memory foundation for all users,” signalling that memory is now treated as a core platform feature rather than an optional add-on. Given the benchmark improvements and the free-tier rollout, that framing seems accurate.