OpenAI began pushing a rebuilt ChatGPT memory system called Dreaming V3 to Plus and Pro subscribers in the United States on June 4, 2026, giving those accounts roughly twice the memory capacity of the system it replaced, per OpenAI’s own release notes. The update replaces the app’s old list of manually saved facts with a background process that reviews a user’s chat history after each conversation ends and updates what the assistant remembers, without anyone typing “remember this.” OpenAI says the new design lifts factual recall accuracy and runs roughly five times more efficiently than the memory system it replaces, a claim echoed across early coverage of the rollout.
The timing is not incidental. ChatGPT’s grip on the AI assistant market has loosened this year as rivals gain ground, and memory, the ability of an assistant to carry context between sessions without being re-told, has become one of the few product features that can lock a user into a single platform. OpenAI has spent 2026 both trimming and expanding its own model lineup, retiring GPT-4.5 from ChatGPT on June 26, 2026 after a 30-day sunset and o3 on August 26, 2026 after a 90-day sunset, then launching GPT-6 globally across ChatGPT’s Free and paid plans in October 2026, even as Dreaming V3 kept shipping on schedule. As of October 2026, OpenAI still lists ChatGPT Plus at $20 a month, now bundled with the expanded memory capacity Dreaming V3 provides. Dreaming V3 is OpenAI’s clearest bet yet that remembering a user well is worth more than any single model upgrade.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What OpenAI Just Shipped: Dreaming V3 in Plain Terms
Dreaming V3 is the name OpenAI gave to a rebuilt version of ChatGPT memory. Instead of a flat list of facts a user explicitly asked the assistant to save, the system now runs a synthesis step after conversations end, reading across a user’s chat history and deciding on its own what is worth carrying forward. OpenAI’s release notes, updated in August 2026, describe memories as now updating automatically, with ChatGPT continuously tracking the details it judges most important rather than waiting on an explicit save command. OpenAI describes three visible changes: a memory summary page that shows what the assistant has pieced together, controls to correct or update any of it, and settings that let a user decide which topics ChatGPT should bring up unprompted and which it should leave alone.
The system also tracks time. A note like “planning a trip to Singapore in July” now gets revised to “traveled to Singapore in July 2026” once the date has passed, rather than sitting in the profile as a stale fact. And when ChatGPT answers using something it remembers, the response shows which past chat, saved detail, custom instruction, or uploaded file the model drew on, giving users a paper trail the original saved-memories list never provided in the same form.
The Rollout Timeline: Who Gets It and When
OpenAI’s own announcement says Dreaming V3 became available on June 4, 2026, for ChatGPT Plus and Pro subscribers in the United States, and the company said at the time that the update would expand from the US to additional countries, and to Free and Go tier users, over the coming weeks, without committing to exact dates. As of publication, OpenAI has not given a specific date for when Team and Enterprise accounts will receive the update.
That staged approach matches how OpenAI has handled every previous memory expansion. The original memory feature launched to a small slice of users in February 2024 before OpenAI confirmed that September it had reached the Free, Plus, Team, and Enterprise tiers, and on April 10, 2025 OpenAI expanded ChatGPT memory so it could reference all past conversations rather than only the facts a user had explicitly saved, letting ChatGPT draw on past-conversation details to personalize web results for all Plus and Pro users except in the EEA, UK, Switzerland, Norway, Iceland, and Liechtenstein. Free-tier users didn’t get that same background-synthesis treatment until June 3, 2025, when OpenAI began rolling out a lighter version built on recent chats; OpenAI described the split at the time as Free accounts getting short-term continuity while Plus and Pro kept the longer-term memory. If Dreaming V3 follows the same pattern, the widest rollout is likely still months away.
Team and Enterprise: Still an Open Question
A workplace deployment raises questions an individual Plus subscriber doesn’t have to answer: who can see what the model has inferred about an employee, whether HR- or legal-sensitive conversations should be excluded from synthesis by default, and how memory is handled once someone leaves the company. Expect OpenAI to treat those questions as a prerequisite for the Enterprise rollout rather than an afterthought.
From “Remember This” to Autonomous Profiles: A Brief History
ChatGPT memory has moved through five distinct phases since OpenAI first introduced it, and the “dreaming” name itself predates the Dreaming V3 headline: OpenAI introduced the first dreaming system in April 2025 to curate memories automatically from chat history, with Dreaming V3 arriving in June 2026 as a more capable successor. Each update handed the assistant a little more initiative to decide, on its own, what was worth keeping.
| Date | Milestone | What Changed | Who Got It |
|---|---|---|---|
| Feb 13, 2024 | Memory launches | Explicit save, view, and forget controls only | Small group of free and paid users |
| Sept 5, 2024 | Wider rollout | Same explicit-save model, broader access | Free, Plus, Team, Enterprise |
| April 2025 | Background synthesis begins | Memory starts referencing all past chats automatically | Plus and Pro first |
| June 3, 2025 | Free tier catches up | Recent-chat referencing added for Free accounts | Free |
| June 4, 2026 | Dreaming V3 launches | Full autonomous synthesis, temporal updates, 82.8% recall | Plus and Pro (US) first |
The direction of travel is consistent: every update shifts more of the remembering from the user to the model itself, culminating in OpenAI making the improved dreaming memory available to Plus and Pro users in the United States in June 2026. Dreaming V3 is the point where that shift becomes close to total. The user no longer curates a list; the model curates an evolving model of the user.
How Dreaming V3 Actually Works Under the Hood
The technical shift is in when and how the model processes memory. Earlier versions leaned on retrieval: the assistant searched a static list of saved facts at the start of a conversation and used whatever matched. Dreaming V3 instead runs its synthesis after a conversation finishes, off to the side, rather than inside the live chat window. OpenAI’s help center added a companion capability in August 2026 called reference chat history, which it says helps ChatGPT locate relevant details from past conversations more reliably even when a user never explicitly flagged them for saving. That separation is why OpenAI can claim a roughly fivefold gain in compute efficiency: the expensive work of reviewing months of chat history happens once, in the background, instead of being repeated at the start of every new session.
The result behaves more like how a person forms an impression of someone over time than like a database lookup. Instead of transcribing what a user says, the system synthesizes, drawing conclusions across many conversations that the user may never have stated outright in any single one of them. That inference step is also where most of the privacy debate below begins.
An Illustrative Example: Static List vs. Synthesized Profile
To see the difference, compare a simplified version of what a saved-memories entry looked like before Dreaming V3 with what a synthesized profile looks like after it. This is an illustrative simplification, not an actual OpenAI API response:
BEFORE — static saved memory (pre-2026):
- "Remember: I'm vegetarian"
- "Remember: I live in Austin, TX"
(unchanged until the user manually edits or deletes it)
AFTER — Dreaming V3 synthesized profile (June 2026):
- Dietary preference: vegetarian (referenced across 14 chats since Jan 2025)
- Location: Austin, TX (last confirmed May 2026)
- Upcoming: trip to Singapore, July 2026
-> auto-revises to "traveled to Singapore" after the date passes
- Confidence: high, drawn from 12 corroborating mentionsThe Numbers: Recall Accuracy and Compute Efficiency
OpenAI has not published a full benchmark table alongside the Dreaming V3 launch, but early coverage of the release converges on one figure: factual recall accuracy of 82.8% under the new system. Reported starting points vary depending on which prior version is used as the baseline, with some accounts citing a jump from roughly 68% and others citing a lower starting point, but the 82.8% endpoint is consistent across multiple independent write-ups of OpenAI’s internal evaluation.
The efficiency claim is more consistent. OpenAI and several outlets covering the release describe Dreaming V3 as about five times more compute-efficient than the memory system it replaces, a gain that comes largely from moving the expensive synthesis work out of the live chat path and into an offline batch process. For a company running one of the highest-traffic consumer AI products in the world, shaving compute cost per active user by that margin is not a minor engineering footnote.
New User Controls: Inside the Memory Summary Page
OpenAI paired the more autonomous backend with a more visible frontend. The memory summary page, reachable from ChatGPT’s settings, lets users review and steer what the system has pieced together, organized by topic, with a timestamp for when each item was last updated. A release-note tracker logged the toggle landing in build ChatGPT v1.2026.209 in August 2026, when a dedicated Memory setting appeared under Settings > Personalization; as of that same OpenAI documentation, memory and reference chat history still ship off by default and must be switched on there before the summary page starts populating. From that page, users can mark an inference as wrong, tell the model to stop bringing up a topic, or remove an item outright.
Inside individual chats, a “don’t mention this again” control lets a user suppress a specific detail on the spot. That traceability addresses a common complaint about the old system, where users often couldn’t tell why ChatGPT knew something about them. It does not fully resolve a separate complaint, covered below, about whether the summary page shows everything the model actually retains.
Why Now? The Business Logic Behind OpenAI’s Memory Bet
Memory is one of the few AI product features that creates real switching costs. A better model is a commodity the moment a competitor matches its benchmark scores, but a chat history that an assistant has spent two years learning from does not transfer if a user switches products. Every one of ChatGPT’s major rivals has reached a similar conclusion. Anthropic opened Claude’s memory to every tier, including free users, on March 2, 2026. xAI shipped persistent memory for Grok in April 2025. Google has taken a different route with Gemini, leaning on a user’s existing Google account data rather than building a parallel chat-history profile from scratch.
Dreaming V3 reads as OpenAI’s answer to that competitive pressure. If a rival’s model briefly pulls ahead on a benchmark, a ChatGPT account that already understands a user’s job, preferences, and ongoing projects gives that user a reason to stay anyway. It is a retention strategy wearing a personalization feature’s clothes.
The Privacy Backlash: Audit Trails and the Consent Question
Not every reaction to Dreaming V3 has been positive. A 2025 survey of 300 US ChatGPT users found that 82% considered their conversations with the assistant sensitive or highly sensitive, a baseline that matters more once the system doing the remembering starts drawing its own conclusions rather than just storing what a user typed. Because Dreaming V3 synthesizes rather than transcribes, the profile it builds can include inferences a user never stated in so many words, assembled from patterns across dozens of conversations.
Researchers have also flagged a narrower, more technical concern: the new architecture’s summary page shows less of the granular audit trail that the old discrete saved-memories list provided. An academic study presented at ACM CHI 2026, titled “Relational Gains, Privacy Strains,” describes this trade-off as a personalization-convenience paradox, the feature users say they value most is also the one they are least able to fully audit or constrain. OpenAI’s response has been to add its own logging, giving security teams a chronological feed of memory creation, edits, and deletions, and to open the Memory API to outside review, inviting researchers from the Berkman Klein Center and the Electronic Frontier Foundation to examine its privacy guarantees.
ChatGPT vs Claude vs Gemini vs Grok: The 2026 Memory Race
With Dreaming V3 live, all four major consumer AI assistants now ship some form of persistent memory, though they differ sharply in default behavior, user control, and portability, according to a side-by-side comparison of the major assistants’ memory systems.
| Assistant | Memory Model | Default Status | User Controls | Cross-App Portability |
|---|---|---|---|---|
| ChatGPT (Dreaming V3) | Autonomous background synthesis of full chat history | On by default for Plus/Pro (US) | Summary page; edit, suppress, or delete items | Stays inside ChatGPT |
| Claude (Anthropic) | 24-hour synthesis cycle; free tier included since Mar 2, 2026 | On, user-editable | Full view/edit/delete in Settings > Capabilities | Stays inside Claude |
| Gemini (Google) | No autonomous chat memory; Personal Intelligence links Google account data | Opt-in, permission-based | Google account-level permissions | Tied to Google account, not chat history |
| Grok (xAI) | Persistent memory since April 2025, plus “Skills” for procedures | On by default where available | View, delete, or disable toggle | Stays inside Grok; excluded EU/UK at launch |
Where the Four Approaches Diverge Most
The practical differences matter more than the marketing language. Claude’s memory is the most transparent of the four, with a dedicated settings panel and a fixed 24-hour synthesis cycle, but it only works inside Claude; switch products and the context resets to zero. Gemini sidesteps building a parallel memory system altogether by drawing on a user’s existing Google account, which is powerful but depends on how much of that account the user is willing to connect. Grok’s rollout has been the most geographically uneven, launching without the EU and UK and only reportedly reaching some accounts in those regions, still in beta, by mid-2026.
Market Impact: Memory as the New Retention Battleground
OpenAI is making this bet while its position at the top of the AI assistant market faces more pressure than at any point since ChatGPT’s launch, with rivals steadily gaining ground. None of that pressure comes from a single source. Google has folded Gemini deeper into Search and Workspace, Anthropic has pushed Claude further into coding and enterprise workflows, and a wave of open-weight models from Chinese labs has made raw model quality look increasingly commoditized.
Against that backdrop, memory is one of the few levers OpenAI controls that a rival cannot simply copy by matching a benchmark score. A user’s two years of context, preferences, and project history inside ChatGPT is not something a competitor can import, and OpenAI’s five-times efficiency claim suggests the company can afford to run that kind of long-term profiling at scale without the compute bill becoming prohibitive. Whether that translates into measurably better retention won’t be clear until OpenAI or an independent analyst publishes usage data tied specifically to Dreaming V3, which has not happened yet.
Regulatory Pressure: GDPR and the EU AI Act Collide With Memory
Dreaming V3’s timing puts it on a collision course with regulation before its global rollout is even finished. The EU AI Act’s transparency requirement, which obligates any AI system to make clear to users that they’re interacting with an AI, becomes mandatory on August 2, 2026 is a transparency-obligation date in the cited materials, not a launch date for Dreaming V3. Separately, under GDPR, an AI system that builds a persistent behavioral profile of a user is generally treated as a profiling activity, which triggers its own consent obligations and gives users a right to request erasure of what has been inferred about them, not just what they typed.
That combination is awkward for a memory system built around inference rather than explicit input. Deleting a fact a user typed is straightforward. Deleting a conclusion the model drew from a pattern across forty conversations, and making sure that conclusion doesn’t quietly reappear the next time the synthesis process runs, is a harder engineering and legal problem. It’s one every AI vendor building autonomous memory will eventually have to answer for European users specifically.
What Comes Next: 5 Predictions for AI Memory
A few things seem likely as Dreaming V3 moves past its first weeks on the market:
- Free and Go tiers catch up within months, not weeks. If OpenAI follows the pattern set by the original memory feature (small group in February 2024, broad access by September), a global expansion beyond Plus and Pro likely lands in late 2026.
- Rivals respond with their own memory upgrades. Expect Anthropic, Google, and xAI to announce or accelerate competing personalization features within the next two quarters rather than cede the retention advantage to OpenAI.
- The EU rollout slows or ships with extra disclosure steps. The August 2 AI Act deadline gives OpenAI a narrow window to adapt Dreaming V3’s consent flow for European users, making a delayed or modified EU release more likely than a simultaneous global launch.
- Independent audits surface specific findings, not just principles. With the Memory API opened to the Berkman Klein Center and EFF, expect published findings, positive or critical, later in 2026.
- Enterprise and Team tiers get governance tools before they get the consumer feature. Admin-level controls over what Dreaming V3 remembers about employees are a harder sell to enterprise buyers than the feature itself, so oversight tooling is likely to ship alongside, rather than after, the Team and Enterprise rollout.
What It Means for Developers and Enterprises
For developers building on OpenAI’s API rather than the consumer ChatGPT app, Dreaming V3 itself is not a direct feature switch; OpenAI has framed the update as a ChatGPT product change rather than an API-wide release. But the architecture matters beyond the consumer app. OpenAI’s decision to publish a Memory API and invite external audits suggests the company expects third-party developers to eventually build on the same synthesis approach, rather than rolling their own memory layer on top of existing API products.
For enterprise buyers, the near-term question is less about the technology and more about governance: who can see what Dreaming V3 has inferred about an employee, how long it persists after that employee leaves the company, and whether it can be excluded from certain conversation types, such as legal, HR, or health, by policy rather than by asking each user to manually opt out. OpenAI’s business release notes, updated in August 2026, describe memory as an adjustable setting that can pull relevant context from past chats, uploaded files, and connected apps to keep an employee’s profile current as their work changes, which raises the stakes for IT teams deciding what to leave connected. Those are the questions IT and security teams are likely to raise before approving a wider rollout, regardless of how good the underlying recall numbers are. A recent academic analysis of ChatGPT’s memory design makes a similar point: the governance model around a memory system matters as much as its technical accuracy.
Frequently Asked Questions
What is ChatGPT Dreaming V3?
Dreaming V3 is OpenAI’s rebuilt ChatGPT memory system, which began rolling out on June 4, 2026. It replaces the old list of manually saved facts with a background process that synthesizes what to remember from a user’s full chat history.
When did Dreaming V3 launch, and who has it?
It launched June 4, 2026, for ChatGPT Plus and Pro subscribers in the United States. OpenAI has said the memory system will expand from the US to additional countries and to Free and Go tier users over the coming weeks, though it has not committed to an exact date.
How is this different from ChatGPT’s original memory feature?
The original feature, launched February 13, 2024, only saved what a user explicitly asked it to remember. Dreaming V3 runs continuously in the background, inferring and updating a profile without requiring an explicit save command.
Can I see and edit what ChatGPT remembers about me?
Yes. A memory summary page in ChatGPT’s settings lists what the system has synthesized, organized by topic, and lets users correct, suppress, or delete individual items.
Is Dreaming V3 available outside the United States?
Not yet at full scale. OpenAI’s initial release covers US Plus and Pro users, with international expansion planned but not dated as of this writing, and any EU release will need to account for the EU AI Act’s August 2, 2026 is described as a transparency-obligation date, not a “transparency requirement” launch tied to Dreaming V3.
How does ChatGPT’s memory compare to Claude, Gemini, and Grok?
Claude’s memory has been available to all users, including free accounts, since March 2, 2026, and runs a 24-hour synthesis cycle. Gemini leans on a connected Google account instead of chat history. Grok has offered persistent memory since April 2025 but launched without EU and UK availability.
Does Dreaming V3 raise privacy concerns?
Yes, and OpenAI has acknowledged as much by opening its Memory API to outside review. Because the system infers details rather than only storing what users type, and because its summary page shows less granular detail than the old saved-memories list, researchers have flagged it as harder to fully audit.
Will Dreaming V3 come to the ChatGPT Free tier?
OpenAI says yes, in the coming weeks after the initial Plus and Pro release, though it has not given an exact date.
