How ChatGPT Memory (Dreaming) Changes AI Personalization for Builders

ChatGPT's memory feature, now powered by an automatic process called dreaming, no longer requires you to explicitly tell it what to remember. Since the June 2024 update, the system continuously builds and refines a profile of your interests, preferences, and context from every interaction Wired. For AI builders, this shift from manual to automatic memory changes how user personalization works, what data gets retained, and how much control users actually have.

What changed with ChatGPT memory

ChatGPT first introduced memory in April 2024, letting users explicitly tell the bot to remember specific facts. That approach was rigid: if you mentioned an upcoming party, ChatGPT would always treat it as future even after the date passed Wired. The June 2024 dreaming update changed that. Memories are now built and edited automatically over time. The system decides what's important enough to remember, and it can update memories as context shifts. OpenAI states three aims for this memory system: carry forward useful context, follow user preferences and constraints, and stay current over time Wired.

Why this matters for AI product teams

If you're building an AI assistant or agent that uses persistent context, ChatGPT's approach is a reference point. Automatic memory means the model can personalize responses without explicit user prompts, which improves convenience but introduces risks. The system can mischaracterize user interests. For example, a week of research on replacing a light switch might lead ChatGPT to treat you as a DIY enthusiast, affecting all subsequent answers Wired. For product teams, this highlights the trade-off between seamless personalization and accurate user modeling. You need to decide how much inference your system should do versus letting users explicitly set preferences.

How to manage memory in practice

Users can control memory through the Personalization settings. In the web app, click your profile icon, then Personalization, then Memory. A toggle switch enables or disables memory. Disabling it stops ChatGPT from using memories but does not delete them Wired. The Memory summaries section shows the AI's profile of you, covering interests and answer delivery preferences. You can highlight any item and choose "Make a correction" or "Don't mention this again." You can also type commands like "forget that I'm a journalist." To wipe everything, use the three-dot menu and select "Delete and turn off memory." Past conversations remain in the system, but if you turn memory back on, ChatGPT will re-index them to rebuild its profile Wired. Temporary chats offer a separate privacy option: nothing in them is committed to memory, and they are forgotten when closed.

Caveats and limitations

The automatic nature of dreaming means the system, not the user, decides what gets remembered. This can lead to mismatches between user intent and remembered data. Privacy considerations are significant: memories may reflect inferred locations, occupations, or interests that users never explicitly shared. OpenAI's documentation may vary with platform updates, so the exact behavior described here may change. For builders, the key takeaway is that memory features require careful design around user consent, data retention, and correction mechanisms.