OpenAI began rolling out Dreaming V3, a new ChatGPT memory architecture, on June 4, 2026, starting with ChatGPT Plus and Pro users in the United States before expanding to more countries and to Free and Go users in the following weeks. The company is selling it as a quality upgrade, but the more important story is architectural: ChatGPT is becoming less like a stateless chatbot and more like a long-running personal computing layer. That shift could make the product dramatically more useful, and also harder to reason about. Memory is where convenience, trust, privacy, and product lock-in all meet.
For most of ChatGPT’s life, memory has been treated as a user-facing convenience. You told the system your writing style, your dietary restrictions, your job role, or the project you were working on, and it tried to carry that forward. Useful, yes, but also somewhat bolted on.
Dreaming V3 suggests a different philosophy. OpenAI is no longer just asking whether ChatGPT can store a few facts about you. It is asking whether the system can continuously synthesize a working model of your preferences, projects, constraints, and current situation across months or years of use.
That is a much bigger bet. A chatbot that remembers your preferred tone is a feature. A chatbot that understands your ongoing work, notices stale assumptions, and adapts its future answers accordingly starts to look like an operating environment for intent.
This matters especially for WindowsForum readers because much of modern computing is already moving away from isolated apps and toward assistants that sit above them. Microsoft has been pushing Copilot across Windows, Microsoft 365, Edge, and enterprise management workflows. OpenAI’s memory work is part of the same race: whoever owns persistent context may own the user relationship.
That version had an obvious strength: visibility. A saved memory list is comprehensible. Users can inspect it, delete entries, and understand why ChatGPT might tailor an answer.
But it also had an obvious weakness. Human beings do not naturally conduct their lives as database administrators. Most users do not stop mid-conversation to say, “Please persist this item for future retrieval.” They mention things casually, change plans, abandon projects, move jobs, switch cities, and develop new preferences without filing a maintenance ticket.
That made saved memory feel powerful in demonstrations and brittle in everyday use. It could remember that you prefer concise answers but miss the fact that your project had changed direction three weeks ago. It could retain a travel plan long after the trip had passed. It could be both too forgetful and too clingy, depending on the moment.
The deeper problem was that saved memory confused two different tasks. Storing a fact is not the same as knowing whether that fact is still useful. A memory system for an AI assistant needs not only retention, but judgment.
The new Dreaming V3 system builds on that idea. Instead of relying primarily on explicit user instructions, it attempts to curate memory as an ongoing background process. In plain English, ChatGPT is being trained not merely to remember what you said, but to maintain a fresher, more useful summary of what those conversations imply.
OpenAI says the new system improves three things: carrying forward useful facts, following user preferences, and staying current as time passes. Its internal evaluations show substantial gains, including factual recall rising from 67.9 percent in the 2025 system to 82.8 percent in the 2026 version. Preference adherence also rises, from 55.3 percent to 71.3 percent, while accuracy over time improves from 52.2 percent to 75.1 percent.
Those numbers should be read carefully. They are internal evaluations, not an independent audit. They also measure success on tasks OpenAI designed or selected, which may not capture the messy edge cases that matter to real users.
Still, the direction is credible. Anyone who uses AI assistants daily understands that long-term usefulness depends less on isolated benchmark cleverness and more on continuity. If ChatGPT can remember the project, the constraints, the user’s preferred format, and the fact that last month’s plan is no longer valid, the product becomes less of a novelty and more of a workstation.
OpenAI says users will be able to review a summary of what ChatGPT knows about them, update details, correct information, and provide guidance about what topics ChatGPT should bring up and when. That is a critical concession to reality. If memory is invisible, users will eventually distrust it.
The old saved-memory model gave users a list. The new model needs something more nuanced, because synthesized memory is not just a pile of stored sentences. It is an interpretation of a user’s history.
That creates a product design challenge. A memory summary must be detailed enough to be useful but not so exhaustive that no one reads it. It must be editable without turning into a second job. It must reveal enough about the assistant’s assumptions that users can correct them before they cause trouble.
This is where OpenAI’s consumer AI ambitions collide with old-school administrative wisdom. Sysadmins know that state is dangerous when it is hidden. Configuration drift, cached credentials, stale group policies, and ghost settings cause problems precisely because systems remember things in places users do not check.
ChatGPT memory is not the same as Windows registry sprawl, but the lesson rhymes. Persistent context is power. Persistent context without inspectability is a support nightmare.
That kind of continuity changes the interaction. Instead of spending the first third of every prompt reconstructing the situation, the user can start closer to the actual task. The assistant can produce answers that are not merely correct in the abstract, but relevant to the user’s situation.
But personalization has a shadow side. The more an assistant adapts to a user, the more its output may become shaped by assumptions the user no longer notices. A model that remembers your preferences can also overfit to them. A model that remembers your constraints can also keep applying old constraints after they stop mattering.
This is not hypothetical. Recommendation systems have spent years teaching the industry what happens when personalization is optimized without enough user agency. The system becomes convenient, then sticky, then subtly narrowing.
In productivity software, the stakes are different from social media, but they are still real. If ChatGPT remembers that you dislike a certain framework, prefer a certain vendor, or work under a certain compliance model, it may steer future answers accordingly. That may save time. It may also quietly reinforce yesterday’s assumptions.
A project manager could ask for a status update and have ChatGPT understand the cadence, stakeholders, and preferred reporting style. A developer could receive code suggestions that reflect the team’s stack. A help desk analyst could draft user-facing explanations in the company’s tone.
Yet enterprise IT will immediately ask harder questions. Where is memory stored? How is it segmented between personal and work contexts? Can administrators disable it? What happens when an employee leaves? Can memory be exported, audited, deleted, retained for legal reasons, or excluded from training?
OpenAI has previously emphasized controls for business plans and has said that Team and Enterprise customer data is not used to train its models by default. But memory adds a new layer to that conversation because it is not merely chat content. It is derived context.
Derived data is often where governance gets messy. A transcript may be deleted, but what about a summary inferred from that transcript? A user may remove a chat, but what if the system already abstracted a preference or project fact from it? These are solvable problems, but they require clear product controls rather than comforting language.
For administrators, the right posture is neither panic nor blind adoption. Treat AI memory like any other persistence mechanism that can contain sensitive operational context. Pilot it, document its behavior, define acceptable-use boundaries, and make sure users understand when to use temporary or memory-disabled modes.
Microsoft’s Copilot strategy depends on context. The assistant is more useful if it knows what document you are editing, what meeting you just attended, what files are relevant, and what enterprise permissions apply. OpenAI’s ChatGPT strategy is converging on the same destination from the web-app side: a persistent assistant that gets better because it knows more about the user over time.
The difference is placement. Microsoft has the operating system, Office graph, identity layer, and management stack. OpenAI has the high-frequency assistant relationship and a large base of users who already treat ChatGPT as a general-purpose workspace.
Dreaming V3 strengthens OpenAI’s position in that contest. If users feel that ChatGPT “knows how I work” better than Copilot does, they may keep returning to it even inside Microsoft-heavy environments. Conversely, Microsoft can argue that memory tied to enterprise identity, permissions, and compliance controls is safer and more manageable.
That competition is good for users only if it produces better controls, not just more aggressive context harvesting. The next phase of AI assistants will not be won merely by bigger models. It will be won by the assistant that can remember enough to be useful, forget enough to be safe, and explain itself enough to be trusted.
Human memory is messy, but it has one advantage: we usually understand that it is fallible. AI memory risks presenting itself with machine confidence even when it is relying on a stale or misinterpreted summary.
Time-sensitive information is especially dangerous. “I am planning a birthday party next Saturday” is useful before the event and misleading afterward. “I am interviewing for a new job” may be relevant for a few weeks and sensitive forever. “My company uses Windows Server 2019” may be correct today and obsolete after a migration.
OpenAI says its new system improves accuracy over time, which is exactly the right metric to target. But users should still expect rough edges. Memory systems will occasionally retain too much, infer too much, or apply context in the wrong setting.
That makes explicit user control essential. The assistant should not merely remember. It should allow users to correct its model of them as easily as they correct a typo.
Memory shifts the debate. The sensitive artifact is no longer just the raw conversation. It is the model’s accumulated understanding of you.
That understanding may include work habits, political interests, health-adjacent details, family structure, financial goals, technical stack, location patterns, and interpersonal relationships. Even if each individual fact seems harmless, the aggregate can be revealing.
This is where OpenAI must be held to a higher standard than ordinary app personalization. A shopping site remembering your shoe size is one thing. A general-purpose AI assistant maintaining a long-term operational profile of your life and work is another.
The memory summary page is a promising answer, but it cannot be the whole answer. Users need clear toggles, temporary modes, deletion semantics, and plain-language explanations of what happens when memory is disabled. Businesses need policy controls and auditability. Regulators will eventually ask whether inferred memory should be treated differently from user-submitted data.
The more useful ChatGPT becomes, the less credible it is to describe memory as a minor convenience. It is becoming a persistent user profile for an AI intermediary.
But memory quality is difficult to benchmark because the failure modes are personal. If ChatGPT forgets your preferred indentation style, that is annoying. If it forgets a medical constraint, a legal instruction, or a security boundary, that is more serious. If it remembers something you wanted confined to one context, that is a different kind of failure.
The most telling test will be whether users start trusting ChatGPT with longer-running workflows. Do writers keep project bibles in memory? Do developers rely on it to remember repository conventions? Do small businesses let it retain operational preferences? Do families use it for planning across months?
If the answer is yes, Dreaming V3 will be remembered as more than an incremental personalization update. It will be a step toward AI assistants as persistent work partners.
If the answer is no, the likely reason will not be lack of raw intelligence. It will be trust. Users can forgive a chatbot for forgetting. They are less forgiving when it remembers wrongly, remembers invisibly, or remembers something it should not.
The practical implications are straightforward, even if the long-term consequences are still developing.
That is why Dreaming V3 matters. It is not just a better memory feature. It is OpenAI’s attempt to make ChatGPT feel less like a website you visit and more like a companion layer across work and life.
For Windows users, sysadmins, and IT pros, the right response is cautious attention. This technology will make some workflows smoother and some governance questions harder. It will save time when it remembers correctly, and it will create risk when users assume that “memory” means perfect, current, and contextually appropriate knowledge.
OpenAI is betting that a more capable memory system will make ChatGPT indispensable. The company may be right. But the future of AI assistants will depend as much on controlled forgetting, user inspection, and administrative discipline as on recall scores, because the assistant that knows you best will also need to prove it deserves to.
OpenAI Is Turning Memory From a Feature Into Infrastructure
For most of ChatGPT’s life, memory has been treated as a user-facing convenience. You told the system your writing style, your dietary restrictions, your job role, or the project you were working on, and it tried to carry that forward. Useful, yes, but also somewhat bolted on.Dreaming V3 suggests a different philosophy. OpenAI is no longer just asking whether ChatGPT can store a few facts about you. It is asking whether the system can continuously synthesize a working model of your preferences, projects, constraints, and current situation across months or years of use.
That is a much bigger bet. A chatbot that remembers your preferred tone is a feature. A chatbot that understands your ongoing work, notices stale assumptions, and adapts its future answers accordingly starts to look like an operating environment for intent.
This matters especially for WindowsForum readers because much of modern computing is already moving away from isolated apps and toward assistants that sit above them. Microsoft has been pushing Copilot across Windows, Microsoft 365, Edge, and enterprise management workflows. OpenAI’s memory work is part of the same race: whoever owns persistent context may own the user relationship.
Saved Memories Were Too Manual for the Job
OpenAI’s first memory system arrived publicly in early 2024 as a controlled rollout and later expanded more broadly. The basic model was easy to understand. Users could ask ChatGPT to remember things, and the system could sometimes infer details worth saving.That version had an obvious strength: visibility. A saved memory list is comprehensible. Users can inspect it, delete entries, and understand why ChatGPT might tailor an answer.
But it also had an obvious weakness. Human beings do not naturally conduct their lives as database administrators. Most users do not stop mid-conversation to say, “Please persist this item for future retrieval.” They mention things casually, change plans, abandon projects, move jobs, switch cities, and develop new preferences without filing a maintenance ticket.
That made saved memory feel powerful in demonstrations and brittle in everyday use. It could remember that you prefer concise answers but miss the fact that your project had changed direction three weeks ago. It could retain a travel plan long after the trip had passed. It could be both too forgetful and too clingy, depending on the moment.
The deeper problem was that saved memory confused two different tasks. Storing a fact is not the same as knowing whether that fact is still useful. A memory system for an AI assistant needs not only retention, but judgment.
Dreaming Is OpenAI’s Answer to Staleness
OpenAI introduced a broader memory approach in 2025, allowing ChatGPT to reference past chat context beyond the explicit saved memories list. Internally, the company described the background process as “dreaming,” a term that is more poetic than precise but captures the intended function: the system reviews prior interactions and synthesizes useful context for future chats.The new Dreaming V3 system builds on that idea. Instead of relying primarily on explicit user instructions, it attempts to curate memory as an ongoing background process. In plain English, ChatGPT is being trained not merely to remember what you said, but to maintain a fresher, more useful summary of what those conversations imply.
OpenAI says the new system improves three things: carrying forward useful facts, following user preferences, and staying current as time passes. Its internal evaluations show substantial gains, including factual recall rising from 67.9 percent in the 2025 system to 82.8 percent in the 2026 version. Preference adherence also rises, from 55.3 percent to 71.3 percent, while accuracy over time improves from 52.2 percent to 75.1 percent.
Those numbers should be read carefully. They are internal evaluations, not an independent audit. They also measure success on tasks OpenAI designed or selected, which may not capture the messy edge cases that matter to real users.
Still, the direction is credible. Anyone who uses AI assistants daily understands that long-term usefulness depends less on isolated benchmark cleverness and more on continuity. If ChatGPT can remember the project, the constraints, the user’s preferred format, and the fact that last month’s plan is no longer valid, the product becomes less of a novelty and more of a workstation.
The Memory Summary Page Is the Real User Interface
The most practically important part of the announcement may not be Dreaming V3 itself. It is the memory summary page.OpenAI says users will be able to review a summary of what ChatGPT knows about them, update details, correct information, and provide guidance about what topics ChatGPT should bring up and when. That is a critical concession to reality. If memory is invisible, users will eventually distrust it.
The old saved-memory model gave users a list. The new model needs something more nuanced, because synthesized memory is not just a pile of stored sentences. It is an interpretation of a user’s history.
That creates a product design challenge. A memory summary must be detailed enough to be useful but not so exhaustive that no one reads it. It must be editable without turning into a second job. It must reveal enough about the assistant’s assumptions that users can correct them before they cause trouble.
This is where OpenAI’s consumer AI ambitions collide with old-school administrative wisdom. Sysadmins know that state is dangerous when it is hidden. Configuration drift, cached credentials, stale group policies, and ghost settings cause problems precisely because systems remember things in places users do not check.
ChatGPT memory is not the same as Windows registry sprawl, but the lesson rhymes. Persistent context is power. Persistent context without inspectability is a support nightmare.
Personalization Cuts Both Ways
The upside is obvious. A ChatGPT that remembers well can spare users from endlessly restating context. It can know that a developer prefers PowerShell examples, that a writer wants AP-style summaries, that a parent is planning around school schedules, or that an IT admin is managing a hybrid Windows 11 and Azure environment.That kind of continuity changes the interaction. Instead of spending the first third of every prompt reconstructing the situation, the user can start closer to the actual task. The assistant can produce answers that are not merely correct in the abstract, but relevant to the user’s situation.
But personalization has a shadow side. The more an assistant adapts to a user, the more its output may become shaped by assumptions the user no longer notices. A model that remembers your preferences can also overfit to them. A model that remembers your constraints can also keep applying old constraints after they stop mattering.
This is not hypothetical. Recommendation systems have spent years teaching the industry what happens when personalization is optimized without enough user agency. The system becomes convenient, then sticky, then subtly narrowing.
In productivity software, the stakes are different from social media, but they are still real. If ChatGPT remembers that you dislike a certain framework, prefer a certain vendor, or work under a certain compliance model, it may steer future answers accordingly. That may save time. It may also quietly reinforce yesterday’s assumptions.
Enterprise IT Will See a Productivity Tool and a Governance Problem
For businesses, long-term AI memory is attractive for the same reason roaming profiles, single sign-on, and cloud-synced settings were attractive. Workers hate repeating themselves. Organizations hate duplicated effort. Persistent context promises smoother onboarding, faster document generation, more consistent analysis, and less time wasted reconstructing background.A project manager could ask for a status update and have ChatGPT understand the cadence, stakeholders, and preferred reporting style. A developer could receive code suggestions that reflect the team’s stack. A help desk analyst could draft user-facing explanations in the company’s tone.
Yet enterprise IT will immediately ask harder questions. Where is memory stored? How is it segmented between personal and work contexts? Can administrators disable it? What happens when an employee leaves? Can memory be exported, audited, deleted, retained for legal reasons, or excluded from training?
OpenAI has previously emphasized controls for business plans and has said that Team and Enterprise customer data is not used to train its models by default. But memory adds a new layer to that conversation because it is not merely chat content. It is derived context.
Derived data is often where governance gets messy. A transcript may be deleted, but what about a summary inferred from that transcript? A user may remove a chat, but what if the system already abstracted a preference or project fact from it? These are solvable problems, but they require clear product controls rather than comforting language.
For administrators, the right posture is neither panic nor blind adoption. Treat AI memory like any other persistence mechanism that can contain sensitive operational context. Pilot it, document its behavior, define acceptable-use boundaries, and make sure users understand when to use temporary or memory-disabled modes.
The Windows Angle Is Not Just Copilot Envy
Windows users should care about this even if they never open ChatGPT in a browser. Memory is becoming one of the central battlegrounds in personal computing.Microsoft’s Copilot strategy depends on context. The assistant is more useful if it knows what document you are editing, what meeting you just attended, what files are relevant, and what enterprise permissions apply. OpenAI’s ChatGPT strategy is converging on the same destination from the web-app side: a persistent assistant that gets better because it knows more about the user over time.
The difference is placement. Microsoft has the operating system, Office graph, identity layer, and management stack. OpenAI has the high-frequency assistant relationship and a large base of users who already treat ChatGPT as a general-purpose workspace.
Dreaming V3 strengthens OpenAI’s position in that contest. If users feel that ChatGPT “knows how I work” better than Copilot does, they may keep returning to it even inside Microsoft-heavy environments. Conversely, Microsoft can argue that memory tied to enterprise identity, permissions, and compliance controls is safer and more manageable.
That competition is good for users only if it produces better controls, not just more aggressive context harvesting. The next phase of AI assistants will not be won merely by bigger models. It will be won by the assistant that can remember enough to be useful, forget enough to be safe, and explain itself enough to be trusted.
Forgetting Is Now a Product Requirement
OpenAI’s announcement focuses on improved recall, but the more interesting challenge is forgetting. A memory system that never forgets becomes polluted. A memory system that forgets too eagerly becomes unreliable. The hard part is deciding which details should decay, which should persist, and which should require explicit confirmation.Human memory is messy, but it has one advantage: we usually understand that it is fallible. AI memory risks presenting itself with machine confidence even when it is relying on a stale or misinterpreted summary.
Time-sensitive information is especially dangerous. “I am planning a birthday party next Saturday” is useful before the event and misleading afterward. “I am interviewing for a new job” may be relevant for a few weeks and sensitive forever. “My company uses Windows Server 2019” may be correct today and obsolete after a migration.
OpenAI says its new system improves accuracy over time, which is exactly the right metric to target. But users should still expect rough edges. Memory systems will occasionally retain too much, infer too much, or apply context in the wrong setting.
That makes explicit user control essential. The assistant should not merely remember. It should allow users to correct its model of them as easily as they correct a typo.
The Privacy Debate Moves From Chat Logs to User Models
Early AI privacy debates focused heavily on prompts and training data. Did the model provider use your chats to improve future models? Could employees see your conversations? Were sensitive files uploaded into a system that should not have them?Memory shifts the debate. The sensitive artifact is no longer just the raw conversation. It is the model’s accumulated understanding of you.
That understanding may include work habits, political interests, health-adjacent details, family structure, financial goals, technical stack, location patterns, and interpersonal relationships. Even if each individual fact seems harmless, the aggregate can be revealing.
This is where OpenAI must be held to a higher standard than ordinary app personalization. A shopping site remembering your shoe size is one thing. A general-purpose AI assistant maintaining a long-term operational profile of your life and work is another.
The memory summary page is a promising answer, but it cannot be the whole answer. Users need clear toggles, temporary modes, deletion semantics, and plain-language explanations of what happens when memory is disabled. Businesses need policy controls and auditability. Regulators will eventually ask whether inferred memory should be treated differently from user-submitted data.
The more useful ChatGPT becomes, the less credible it is to describe memory as a minor convenience. It is becoming a persistent user profile for an AI intermediary.
Benchmarks Are Encouraging, but Daily Use Will Decide
OpenAI’s reported gains are substantial. A jump from 67.9 percent to 82.8 percent in factual recall is not cosmetic. Preference adherence moving from 55.3 percent to 71.3 percent could noticeably improve everyday use.But memory quality is difficult to benchmark because the failure modes are personal. If ChatGPT forgets your preferred indentation style, that is annoying. If it forgets a medical constraint, a legal instruction, or a security boundary, that is more serious. If it remembers something you wanted confined to one context, that is a different kind of failure.
The most telling test will be whether users start trusting ChatGPT with longer-running workflows. Do writers keep project bibles in memory? Do developers rely on it to remember repository conventions? Do small businesses let it retain operational preferences? Do families use it for planning across months?
If the answer is yes, Dreaming V3 will be remembered as more than an incremental personalization update. It will be a step toward AI assistants as persistent work partners.
If the answer is no, the likely reason will not be lack of raw intelligence. It will be trust. Users can forgive a chatbot for forgetting. They are less forgiving when it remembers wrongly, remembers invisibly, or remembers something it should not.
The Practical Shape of This Upgrade Is Already Visible
OpenAI’s announcement gives users and administrators enough concrete detail to start forming a plan. This is not a speculative lab feature; it is rolling out now to paying users in the United States and is expected to expand more broadly.The practical implications are straightforward, even if the long-term consequences are still developing.
- ChatGPT Plus and Pro users in the United States are first in line for the Dreaming V3 memory upgrade, with broader availability planned over the following weeks.
- The new system is designed to synthesize memory from past conversations more efficiently, not merely expand the old list of saved memories.
- OpenAI’s internal evaluations show better factual recall, stronger preference following, and improved handling of time-sensitive context compared with prior memory systems.
- The memory summary page is the control surface users should watch, because it determines whether synthesized memory feels transparent or creepy.
- Free and Go users are expected to receive the improved memory system later, helped by OpenAI’s claimed compute-efficiency gains.
- IT teams should treat AI memory as persistent derived context and decide where it is appropriate before users make it part of routine workflows.
OpenAI’s Real Product Is Continuity
The larger lesson is that AI assistants are moving past the era of impressive one-off answers. The next competitive frontier is continuity: the ability to understand what came before, adapt to what changed, and help without forcing the user to rebuild context every time.That is why Dreaming V3 matters. It is not just a better memory feature. It is OpenAI’s attempt to make ChatGPT feel less like a website you visit and more like a companion layer across work and life.
For Windows users, sysadmins, and IT pros, the right response is cautious attention. This technology will make some workflows smoother and some governance questions harder. It will save time when it remembers correctly, and it will create risk when users assume that “memory” means perfect, current, and contextually appropriate knowledge.
OpenAI is betting that a more capable memory system will make ChatGPT indispensable. The company may be right. But the future of AI assistants will depend as much on controlled forgetting, user inspection, and administrative discipline as on recall scores, because the assistant that knows you best will also need to prove it deserves to.
References
- Primary source: Neowin
Published: 2026-06-05T05:22:07.410753
OpenAI is rolling out a major upgrade to ChatGPT memory
Built on top of the "dreaming" background process introduced in 2025, ChatGPT's new compute-efficient memory update can carry forward useful context for users.
www.neowin.net
- Official source: openai.com
Dreaming: Better memory for a more helpful ChatGPT
ChatGPT introduces a new memory system to better remember preferences, keeping context fresh and relevant across conversations.openai.com
- Official source: help.openai.com
ChatGPT — Release Notes | OpenAI Help Center
A changelog of the latest updates and release notes for ChatGPT
help.openai.com
- Related coverage: techtimes.com
OpenAI Improves ChatGPT Memory Feature on Free Tier With New 'Dreaming' Architecture
OpenAI unveils ChatGPT’s latest “dreaming” architecture upgrade on its Memory feature for free accounts. Here’s what the upgrade is all about.
www.techtimes.com
- Related coverage: techradar.com
- Related coverage: m.investing.com
Dreaming memory system rolls out to ChatGPT Plus and Pro users By Investing.com
Dreaming memory system rolls out to ChatGPT Plus and Pro usersm.investing.com
- Related coverage: macrumors.com
ChatGPT Gets Better at Remembering Info About You With Memory Improvements
OpenAI today updated ChatGPT with an improved memory feature that allows the chatbot to remember more about your conversations. OpenAI CEO Sam Altman says that ChatGPT can "reference all your past conversations," allowing for a more personalized experience. ChatGPT will remember context from...
www.macrumors.com
- Related coverage: tomsguide.com
- Related coverage: unanswered.io
ChatGPT Memory Explained: Limits, Types & Control
Discover how ChatGPT's memory stores 1,200-1,400 words of details plus infinite chat history. Learn what it remembers, capacity limits, and how to manage it for better personalization.unanswered.io - Related coverage: jobirun.com
［AIツール利用者向け］ChatGPTのメモリが「夢で学習」する仕組みに進化——Dreaming V3が示す記憶の刷新 | JOBIRUN
はじめに OpenAIが2026年6月4日、ChatGPTのメモリシステムを大幅に刷新したと発表しました。「Dreaming」と呼ばれるバックグラウンド処理で会話履歴を自動統合する新アーキテクチャにより、事実の記憶・嗜好への適応・時間経過
jobirun.com
- Related coverage: 9to5mac.com
OpenAI says ChatGPT's memory feature is getting smarter and coming to free users - 9to5Mac
ChatGPT’s memory feature allows it to learn about you and remember your preferences over time. Memory gives ChatGPT useful context...
9to5mac.com
