Introduction
OpenAI CEO Sam Altman has once again sketched out an ambitious version of the future AI assistant.
In a recent conversation with Silicon Valley interns, Altman described a system that could continuously understand what is happening on your computer, follow the meetings and calls you choose to share, and build enough context to understand the broader flow of your digital life.
His estimate was striking: this kind of assistant may be only one model generation away from becoming genuinely useful, with a rough timeline of around six months.
That is more than a simple memory upgrade or a larger context window. The vision is of an AI that is no longer present only when you open a chat. Instead, it stays contextually aware, understands what you have been doing, and can help without forcing you to repeatedly explain the same people, projects, files, decisions, and background.
If that direction becomes practical, the relationship between people and AI could change significantly: from occasional conversations to persistent assistance, and from isolated tasks to a continuously accumulated personal context.
The more important question is whether Altman’s timeline is realistic—and whether users will actually want an AI system to know that much.
Altman: Let AI Remember the Context of Your Whole Life
Altman has been talking about extremely long-term AI memory for some time.
One version of the idea is surprisingly simple: a small, efficient reasoning model connected to an enormous store of personal context, potentially spanning years of conversations, files, preferences, projects, and activity. New information would keep being added instead of every interaction beginning from scratch.
At Sequoia’s AI Ascent event, Altman also described how different age groups already use ChatGPT in noticeably different ways.
His rough breakdown was:
- Older users often treat ChatGPT as a replacement for search.
- People in their twenties and thirties are more likely to use it as a life adviser.
- College students increasingly treat it more like an operating system for their work and decisions. The trend becomes more interesting with younger users. Altman noted that some people increasingly consult ChatGPT before making important life decisions because the system already has context about people, relationships, projects, and previous discussions.
That behavior hints at what a more persistent AI could become.
This time, however, Altman’s description went beyond conversational memory. He talked about an AI that could understand your screen activity, meetings, calls, texts, email, documents, Slack, and other sources you explicitly connect.
The goal is not necessarily for AI to make every decision on your behalf. A more plausible role is a continuously informed collaborator: something that knows what you are working on, remembers what happened previously, drafts the next step, notices inconsistencies, and helps you resume work without reconstructing the entire context.
That would be a meaningful shift from “use AI when needed” to “AI is continuously available with relevant context.”
The more of your work and personal history you place into such a system, however, the more valuable—and harder to replace—that system becomes.
Memory Is Moving From Passive Storage to Active Synthesis
Some pieces of this future are already visible in current products.
ChatGPT can use memory from past conversations, and its voice experiences can accept screen context in supported environments. ChatGPT Record can also transcribe and summarize meetings, brainstorms, and voice notes on supported platforms.
On the coding side, OpenAI previously introduced an experimental Codex feature called Chronicle.
Chronicle was designed to give Codex recent memory of what a user had been doing on their computer. It used screen context to help Codex understand references such as “this,” “that,” or a project the user had been working on earlier, reducing the need to repeat background information.
Greg Brockman described the experience as “surprisingly magical.”
Chronicle was never the same thing as a complete “memory of your entire life.” It was an opt-in research preview focused on recent computer context for Codex workflows.
It also came with meaningful security considerations. OpenAI’s documentation warned that screenshot-derived context could increase exposure to prompt injection, and Chronicle stored its generated memories locally on the device. The feature was user-controlled and could be paused.
Chronicle Has Now Evolved Into Computer History
There is an important update to the original article.
As of August 14, 2026, OpenAI’s current documentation says Computer History replaces the earlier Chronicle research preview, although OpenAI describes it as a rebuilt system rather than a simple rename.
The difference is significant.
Chronicle used screenshots. Computer History instead records interaction events from apps and websites that the user allows, such as clicks, typing, keyboard shortcuts, app switches, and context exposed through macOS accessibility APIs. It periodically turns those events into text summaries and local memory files.
OpenAI states that the current Computer History system does not capture screenshots, screen recordings, microphone input, or system audio.
It remains opt-in, and users can control which apps and websites contribute, pause collection, inspect their history, and delete stored items.
That evolution shows how quickly this area is changing. The high-level goal remains the same—giving AI durable context about recent work—but the implementation is already shifting toward more explicit controls and a narrower data-capture model.
Dreaming Makes ChatGPT Memory More Dynamic
The other major development is Dreaming, OpenAI’s newer memory architecture for ChatGPT.
Dreaming moves beyond a static list of facts. Instead, ChatGPT can synthesize information from conversation history into a memory state and update that state as time passes.
A simple example explains the difference.
Suppose you tell ChatGPT, “I’m going to Singapore in July.”
A static memory system may continue treating that trip as upcoming even after July has passed. Dreaming is designed to understand the passage of time and revise the memory so that the system can later interpret it as a past trip rather than a future plan.
That sounds minor, but it solves a real problem with persistent assistants: old memories can become misleading if they are never updated.
OpenAI says the system is better at recalling relevant facts, following user preferences, and staying accurate when time changes the meaning of remembered information.
The source article highlighted internal evaluation figures showing improvements across factual recall, preference adherence, and time-sensitive accuracy. One of the clearest published comparisons shows time-sensitive task accuracy rising from 52.2% in 2025 to 75.1% in 2026.
OpenAI also says recent improvements reduced the compute required to serve Dreaming to free users by roughly 5×, making broader rollout practical.
The key idea is that memory is becoming less like a notebook and more like a maintained model of the user’s current situation.
Users still retain controls. OpenAI provides a memory summary that can be reviewed and edited, and memory can be disabled. Temporary Chat can also be used for conversations that should not contribute to persistent memory.
Dreaming and Computer History are different systems, but together they point in the direction Altman described:
- Dreaming builds longer-term personalized context from conversations.
- Computer History adds recent activity context from the user’s computer.
- Record mode can capture and summarize meetings or spoken discussions.
- Screen context can help ChatGPT understand what the user is currently looking at in supported experiences. None of these individually equals a complete “AI that knows your whole life.” Together, however, they form recognizable pieces of that architecture. Memory Is Becoming a New Competitive Moat OpenAI is not alone in treating persistent memory as a major product layer. Almost every major AI platform is moving in the same direction. Google: Memory as Agent Infrastructure Google’s approach is strongly tied to enterprise agents. Its Agent Platform Memory Bank creates long-term memories from conversations between a user and an agent. These memories can persist across sessions, allowing an agent to recall preferences, important facts, and previous context without keeping the entire conversation history inside the active model context window. Google positions Memory Bank as infrastructure for developers building stateful and long-running agents. That makes it slightly different from a consumer-facing ChatGPT memory feature, but the underlying goal is similar: an agent should not forget everything when one session ends. Anthropic: Persistent Memory Across Claude Anthropic has also expanded memory across Claude. In March 2026, memory from chat history became available to all Claude users, including free users. Claude can use previous conversations to create continuity across chats, while users can manage memory and use privacy-oriented options such as Incognito chats. Anthropic also provides a memory tool for developers building agent applications. The tool allows Claude to create, read, update, and delete persistent memory files across sessions. The practical result is the same competitive pressure: assistants are expected to remember the user, not merely answer the latest prompt. Mem0: A Cross-Platform Memory Layer Third-party projects are taking a different approach. Mem0 positions itself as a universal memory layer for AI assistants and agents. Instead of tying memory to a single model provider, developers can use an independent memory layer around different AI systems. That matters because platform-specific memory creates an obvious portability problem. If your most useful context lives inside ChatGPT, moving to Claude may mean starting over. If you use ChatGPT, Claude, and Gemini in parallel, each system may gradually build a different representation of you based on the conversations and data it has seen. A cross-platform memory layer tries to separate the user’s persistent context from the model itself.
The Hidden Lock-In of AI Memory
Model quality changes quickly.
A model that leads benchmarks today may be overtaken several months later. Users can often switch between models with relatively little effort if all they care about is raw reasoning or coding quality.
Memory is different.
The value of memory compounds over time. A system that already understands your preferences, work habits, recurring projects, important relationships, writing style, tools, and previous decisions has an advantage that cannot be reproduced instantly by switching to another model.
That creates a subtler form of platform lock-in.
The real switching cost may eventually be less about losing access to a particular model and more about losing years of accumulated context.
This is why portability, export controls, privacy settings, and the ability to delete or isolate memories will become increasingly important.
It also explains why Altman’s vision is not only a product prediction. It has major business implications.
The company that becomes the place where a user’s long-term AI memory lives may gain a relationship that is far more durable than one built on model performance alone.
At the same time, the more complete that memory becomes, the higher the stakes for privacy and security. A system that knows what you are doing can be extremely helpful, but it also needs clear opt-in controls, strong boundaries, transparent storage behavior, and an easy way for users to inspect and remove what has been remembered.
The future “ultimate AI assistant” will therefore depend on more than better models.
It will also depend on whether users trust the memory layer.
FAQ
What did Sam Altman say about the future of ChatGPT?
Altman described a future AI assistant that could understand a user’s screen activity, meetings, calls, messages, documents, and other connected context. He suggested that making this genuinely useful may require only one more model generation, with a rough timeline of around six months.
Does ChatGPT currently watch my screen all the time?
No. Current screen-context and computer-history features are user-controlled and limited to supported environments. OpenAI’s current Computer History documentation says the feature is opt-in and does not record screenshots, screen video, microphone input, or system audio.
What was OpenAI Chronicle?
Chronicle was an opt-in Codex research preview that built memories from recent screen context. OpenAI now says Computer History has replaced Chronicle with a rebuilt system based on interaction events rather than screenshots.
What is ChatGPT Dreaming?
Dreaming is OpenAI’s newer memory architecture for ChatGPT. It synthesizes useful context from previous conversations and can update memories over time so temporary facts—such as future travel plans—do not remain permanently stale.
Can I turn ChatGPT memory off or edit what it remembers?
Yes. OpenAI provides controls for reviewing, editing, deleting, and disabling memory. Temporary Chat is also available when you want a conversation that does not use or update persistent memory.
What is Google Memory Bank used for?
Google’s Agent Platform Memory Bank gives AI agents persistent, long-term memory across multiple sessions. It is designed primarily as infrastructure for developers building personalized and stateful agents.
Does Claude have persistent memory?
Yes. Anthropic has memory from chat history in the Claude app and also offers a memory tool for developers building agent systems. These features are designed to preserve relevant context across separate conversations.
Why could AI memory create platform lock-in?
Long-term memory becomes more valuable the longer a user stays with a platform. Switching to another assistant can mean losing accumulated preferences, history, project context, and personalization unless memory can be exported or shared across systems.
Top comments (0)