ChatGPT Just got Advanced Memory and it's Creepy... but SO COOL!

ChatGPT’s new Memory feature is rolling out to a limited slice of free and Plus users, letting the assistant remember personal details and preferences across chats—making responses feel more tailored while also introducing new privacy and “creepiness” concerns. Users can explicitly tell ChatGPT what to remember (for example, “I like concise responses” or “I have a 2-year-old dog named Oscar”), and the system can also update what it remembers over time—such as replacing an outdated note (“waiting for new glasses”) with the updated reality (“has new glasses”). The feature works with both GPT 4 and GPT 3.5, and it includes a “temporary chat” mode that behaves like incognito: it won’t appear in history, won’t use Memory, and won’t be used to train the model.

In practice, Memory management is done through a dedicated memory area where users can review, clear, and update stored items. The transcript shows quick confirmation when a memory is added (a “memory updated” icon), plus the ability to instruct the assistant to forget specific information. However, the rollout feels uneven: updates can be delayed, and sometimes new information appears to be stored without an obvious “memory updated” notification. That inconsistency becomes a key tension—users may not always know when the system is learning, which matters when the assistant starts inferring details from context.

A major demonstration centers on the assistant learning about the user’s YouTube channel. After being prompted to infer details from a channel screenshot and later encouraged to do web research, the assistant begins storing structured facts such as subscriber count, video count, themes, and even personal attributes it hadn’t been explicitly told. The transcript also describes a “creepy” moment where the assistant appears to follow instructions embedded in the channel’s “about” section—delivering a greeting and a subscription prompt as part of its stored knowledge. The result: Memory can become a form of automation, where future chats start with the assistant already aware of the user’s context and goals.

The Memory feature is positioned as a productivity upgrade for creators and frequent users. Once the assistant knows someone’s role and typical tasks, it can generate more relevant outputs—like writing YouTube descriptions tailored to the user’s audience even when the user doesn’t provide full context in the prompt. Still, the transcript raises practical questions: how large Memory can grow before it hits model limits, whether Memory will carry into Custom GPTs, and how much control users will have—especially on mobile.

Finally, the transcript notes that Memory management appears to be desktop-only for reviewing and editing stored items, while the mobile app can still create memories without showing when they’re added or removed. The overall takeaway is a tradeoff: Memory makes ChatGPT more useful and context-aware, but it also increases the need for transparency, user control, and careful consideration of what personal information gets stored—particularly as the feature inches toward more autonomous, life-managing behavior.