OpenAI has quietly reshaped how ChatGPT handles personal data on Android, and the change is smaller in scope but bigger in implication than most software updates get credit for. The company’s latest ChatGPT personalization update, nicknamed the “Sunspot refresh,” gives users the ability to control memory and chat history one conversation at a time, rather than flipping a single account-wide switch. It’s currently rolling out only to the Android beta channel, but it marks a meaningful shift in how OpenAI thinks about user data.
Summary
Key takeaways
- OpenAI’s Sunspot refresh adds a per-chat memory toggle and a save-to-history option for ChatGPT’s Android beta app.
- Users can now decide whether the AI remembers a given conversation, and separately, whether that conversation gets stored in their chat history at all.
- The update follows a multi-year evolution of ChatGPT memory, starting with a blunt on/off account-wide setting introduced on February 13, 2024.
- An August 14, 2026 release added project-only memory before Sunspot extended similar granularity down to individual chats.
- The features are limited to the Android beta channel for now, with no confirmed timeline for a broader rollout.
What the Sunspot refresh actually changes
The Sunspot refresh introduces two controls that sit at the conversation level rather than the account level. The per-chat memory toggle lets a user turn memory on or off for one specific chat, without touching how ChatGPT behaves everywhere else. The save-to-history option works alongside it, letting people decide whether a particular exchange gets stored in their history at all, or disappears once the session ends.
Together, these two controls give ChatGPT users something they’ve been asking for: the ability to treat sensitive or one-off conversations differently from the ones they want the AI to remember long-term. That distinction didn’t exist in earlier versions of the app, where memory settings applied uniformly across every chat a user had.
Why granular controls matter more than they sound
On paper, a toggle sounds like a minor interface tweak. In practice, per-chat controls change the calculus for anyone using ChatGPT for something they’d rather not have permanently logged, whether that’s drafting a sensitive email, testing an idea, or asking a question they don’t want influencing future responses. This kind of granularity is exactly what privacy regulators, particularly in the European Union, have been pushing AI companies to offer.
The long road to granular memory in ChatGPT
This ChatGPT memory controls update didn’t appear out of nowhere. OpenAI first introduced memory and user-level controls back on February 13, 2024, and that early version was about as simple as it gets: memory was either on for the entire account or off entirely, with no room for exceptions.
The system got more sophisticated from there. By April 2025, OpenAI rolled out an upgrade that let certain users have ChatGPT reference their chat history directly, giving the model a broader sense of context across separate conversations rather than treating each one in isolation.
By 2026, the architecture had grown into something considerably more layered. ChatGPT’s memory system by that point combined saved memories, the ability to reference chat history, and temporary chats that left no trace, three distinct modes operating side by side. Then, on August 14, 2026, OpenAI pushed the concept further with project-only memory, letting the AI retain context within a specific project without that information bleeding into unrelated conversations.
The Sunspot refresh is the next logical step in that progression, applying the same kind of compartmentalization OpenAI built for projects down to the level of a single chat. Notably, the timing lines up with what OpenAI described as a wave of user requests in mid-2026 specifically asking for this kind of per-conversation control.
Why this fits a broader privacy push
Regulatory pressure, especially from EU privacy law, has been steadily pushing AI companies toward giving users more precise say over how their data gets used and stored. Blanket memory settings satisfy the letter of consent requirements but not necessarily the spirit of them, since a single toggle can’t account for the difference between a routine question and a deeply personal one.
Per-conversation controls represent the kind of granularity regulators have been asking for, and they also happen to align with what users have been requesting directly. That overlap between regulatory expectation and user demand is likely why OpenAI’s OpenAI privacy features have moved in this direction rather than toward some other kind of personalization tool.
Whether other AI chatbots follow suit with similarly granular controls remains an open question, but OpenAI’s incremental rollout, from account-wide toggles in 2024 to project-only memory in mid-2026 to per-chat controls now, suggests this won’t be the last adjustment. For now, the Sunspot refresh ChatGPT update is confined to Android beta testers, leaving the rest of ChatGPT’s user base watching to see when, or if, the same controls reach their own apps.
FAQ
What new features does the Sunspot refresh add to ChatGPT’s Android beta?
It adds a per-chat memory toggle allowing users to enable or disable memory for individual conversations, along with an option to save individual conversations to history.
Why is per-chat memory control important?
Per-chat controls offer granular data management that aligns with privacy law requirements and reflects user demand for controlling exactly what the AI remembers from a given conversation.
Is the Sunspot refresh available on all devices?
No. The Sunspot refresh features are currently limited to the ChatGPT Android beta channel.
How has ChatGPT’s memory system evolved prior to this update?
It started as an account-wide, binary on/off setting in February 2024, then expanded to referencing chat history in April 2025, grew into a multi-layered memory system by 2026, and added project-specific memory capabilities in August 2026 before this per-chat update arrived.
Article produced with the assistance of artificial intelligence and reviewed by the editorial team.
