ChatGPT Has Gained “Living Memory”: The Assistant Becomes More Useful, but Someone Else’s Version of You Can Persist
How does ChatGPT's “living memory” make the assistant more useful while also entrenching an inaccurate version of the user?
Break down the price of personalization in ChatGPT and Gemini and return controllable authority to the user; the assessment must check the scope of data and permissions, retention rules, and the ability to revoke access.
What to watch for
Key takeaways
The decision in “What OpenAI released: transition to a ‘living memory’” depends on one criterion: “living memory” means the system gathers context across conversations and uses it without a separate request, whereas the old ChatGPT started almost from zero every time.
For the “Memory & Dreaming: 3 problems OpenAI wants to solve” scene, the decisive point is this: memory is not one list of facts — there is current context, stored information, project history, work sources, and a summary of behavior — and OpenAI is trying to connect these layers into a single profile.
The “What ‘memory’ in AI is: 5 levels of AI memory” topic becomes clearer once this point is included: there are several memory layers, and Gemini and Copilot already use them, so it is useful to ask which layer drives an answer and whether you can see it.
In the context of “Dreaming v3 from OpenAI,” this criterion applies: Dreaming periodically reprocesses fragments — merging repetition, keeping what matters, and removing noise — but there is no guarantee it keeps what the user would have chosen.
In the context of “How Dreaming v3 differs from old memory,” this criterion applies: unlike passive storage, Dreaming summarizes behavior, which helps but also risks fixing a wrong conclusion that then invisibly shapes new answers.
The working conclusion from “Main risk of new memory” is that a model may once decide that a person works in another industry, prefers a particular approach, or is connected to a particular project. The false conclusion then begins affecting new answers. The user sees the result, but not always the memory that formed it.
The decision in “How new memory affects the AI market and the human being” depends on one criterion: memory therefore has to be visible, editable, and separated by project. An assistant really can remember you—but the right to determine exactly what it remembers should remain yours.
What this episode is about
OpenAI is developing Memory and Dreaming so the model can connect conversations, projects, and habits. This moves closer to true personal AI, but creates a new risk: the system can retain a mistaken conclusion, mix contexts, and influence future answers without showing the user where the idea came from.
The old ChatGPT began almost from zero every time. A person had to explain their role, project, style, and constraints again. “Living memory” changes that relationship: the system gathers context across conversations and tries to use it without a separate request.
Memory is not one list of facts. There is current context, stored information, project history, work sources, and a deeper summary of behavior. Gemini and Microsoft Copilot already use several layers, while OpenAI is trying to connect them into a single profile through Memory and Dreaming.
Dreaming is meant to prevent the model from merely accumulating thousands of fragments. It periodically reprocesses them, combining repetition, identifying what matters, and removing noise. This resembles human generalization from experience, but the algorithm offers no guarantee that it will consider important what the user would have chosen.
The main risk is a persistent error. A model may once decide that a person works in another industry, prefers a particular approach, or is connected to a particular project. The false conclusion then begins affecting new answers. The user sees the result, but not always the memory that formed it.
Personal AI will change the market more than another benchmark increase because accumulated context keeps a person inside a platform. Memory therefore has to be visible, editable, and separated by project. An assistant really can remember you—but the right to determine exactly what it remembers should remain yours.
Deep context improves the answer while raising the cost of error: the product must explain what is stored, who has access, and how to leave without losing history.
Episode transcript
The episode is in Russian; below is an English reading guide to the transcript (the full EN transcript is a machine translation). Voice matching applied to 2 segments: 2 identified, 0 mixed, 0 probable, and 0 unresolved.
Loading…