OpenAI CEO Sam Altman has once again stunned the tech world.
In a recent conversation with Silicon Valley interns, he bluntly stated that the "ultimate AI assistant" is just "one model iteration / 6 months" away.
In the next six months, we are about to enter a world where successors to ChatGPT can watch your screen, record every meeting and call, and have a perfect grasp of the entire context of your life.
We are just one generation of model iteration away from this functionality becoming extremely practical.
Note, he didn't say "stronger memory" or "larger context window", but rather continuous screen watching, recording every meeting and call, and mastering the perfect context of your entire life.
This will be a qualitative leap in human-computer relationships: from conversation to symbiosis. But this time, is Altman's prophecy reliable?
Altman: Let AI Remember All of Your Life
Over the past two years, Altman has repeatedly envisioned a goal: a very small reasoning model with a trillion-token context window that can remember all of your life and continuously add to it.
At the Sequoia AI Ascent conference last May, Altman offered a rough but compelling observation.
He made a broad yet clear division:
Older generations use ChatGPT as a replacement for Google;
People aged 20 to 30 use it as a life coach;
College students treat it directly as an operating system.
The further trend is that young people "consult ChatGPT first for life's major decisions".
These are all signals of the future new form of AI. But this time, he's not just talking about "memory"; he directly names screen perception and comprehensive meeting recording.
Altman boldly claims: Within the next 6 months, the next generation of AI will completely understand you.
This will be a true leap, from "occasional use" to "continuous presence," from "single tasks" to "life-level context accumulation." The more of your life you invest in this AI, the greater the cost of leaving it.
Memory: From Passive Storage to Active Synthesis
The prototype has already begun to emerge: the ChatGPT desktop app can already access screen context in voice mode, and the Enterprise version has long had meeting recording and summarization features.
On April 21 this year, OpenAI launched a research preview feature called Chronicle in its programming product Codex.
It works by taking intermittent background screenshots + OCR, generating memories about "what you've been doing recently," saved as locally encrypted Markdown files. This helps Codex understand "that project" from two weeks ago without the user needing to repeat themselves.
After trying it out, OpenAI Chairman Greg Brockman used two words: "surprisingly magical."
Its internal codename is "telepathy". Altman himself said, "It lives up to the name, that's exactly what it feels like."
But Chronicle ≠ comprehensive "life context." It mainly serves the Codex programming workflow, capturing recent screen content to supplement context, rather than continuously recording meetings/calls or building a complete personal profile. Users can pause it anytime.
Moreover, the risks are higher. Official documentation explicitly warns about prompt injection, unencrypted local storage, and requiring user authorization.
Another move by OpenAI is the launch of Dreaming V3 for ChatGPT in June this year.
Dreaming V3 is the latest architecture for ChatGPT's memory system: it automatically synthesizes memory states from conversation history in the background, has temporal awareness, reduces outdated information, and improves factual recall and preference adherence.
Users can view/edit memories via a memory summary page.
Most interestingly, it has automatic evolution capability. If you said "going to Singapore in July," once that time passes, the system will automatically change the memory to "you went to Singapore in July." Memory is no longer static; it evolves with time.
In OpenAI's internal evaluation, this update showed significant performance improvements:
Factual recall rate increased from 67.9% in 2025 to 82.8%;
Preference adherence increased from 55.3% to 71.3%;
Temporal sensitivity accuracy increased from 52.2% to 75.1%.
Furthermore, computational efficiency improved by about 5 times, which is also why it could be rolled out to the free user tier.
Dreaming V3 and Chronicle are parallel but separate product line advancements—one focuses on screen context (Codex programming scenarios), the other on conversational synthesis (ChatGPT general use).
Together, they move closer to the "always-on, deeply personalized" direction Altman described, but are still in the research preview/phased rollout stage.
Memory is Becoming the New Moat
Altman's vision isn't a solo effort. Zooming out, it's clear that in the past few months, almost the entire AI industry has bet on the same direction at the same time.
Google is taking the enterprise route.
Memory Bank launched on the Gemini enterprise Agent platform. Developers can equip their own AI agents with a cross-session memory pool capability, automatically extracting user preferences, key milestones, and explicit instructions from conversations.
This line is currently more focused on developer infrastructure and hasn't reached the ordinary consumer's Gemini chat dialog yet, but the direction aligns with OpenAI's.
Anthropic isn't idle either.
In March this year, persistent memory was rolled out to all Claude users, both free and paid versions, with the background quietly summarizing data from conversations.
In April, a public beta of persistent memory was added to the developer-side Managed Agents. Enterprise clients like Netflix are already using it in production, with some reporting a 97% reduction in errors.
Third parties are also entering the fray, but with the opposite approach.
Projects like Mem0 aim to become a cross-platform memory middleware layer, interfacing simultaneously with OpenAI, Anthropic, and open-source models, preventing developers from being locked into any single provider.
This precisely indicates the industry's awareness of the problem—if every giant builds its own memory silo, the ones who suffer in the end will be the users and developers trapped within them.
Right now, almost all major players are sprinting in this direction.
But what no one wants to openly admit is that your AI memory is now locked into the platform you use most. Switching from ChatGPT to Claude means an empty memory, starting from scratch. Using three tools simultaneously means you're cultivating three separate, non-communicating, and differently evolving versions of "you."
This is a more hidden moat than model performance.
Model strengths can flip rankings in three months, but memory becomes thicker the longer it accumulates. The thicker the accumulation, the higher the cost of leaving. Whoever accumulates users' memories thick enough first effectively welds the users to their platform.
Understanding this layer reveals that Altman's statement isn't just a product vision; it's also a business prediction.
References:
https://x.com/OpenAIDevs/status/2046288243768082699
https://x.com/haider1/status/2087219995466224120
https://x.com/cory/status/2087060650870907170
https://www.youtube.com/watch?v=gXsutRiJbZI
https://inferencebysequoia.substack.com/p/openais-sam-altman-on-building-the?utm_source=publication-search
This article is from the WeChat public account "New Zhiyuan", author: ASI Revelation
