Everything reminds me of Her. While ChatGPT is not as powerful as the artificial intelligence from Spike Jonze’s sci-fi romance movie, OpenAI’s experimental memory tool for its chatbot seems to suggest a future where bots are highly personalized and capable of more fluid, lifelike conversations.
OpenAI just soft-launched a new feature for ChatGPT called Memory, where the AI chatbot stores personal details that you share in conversations and refers to this information during future chats. Right now, ChatGPT’s Memory feature is available only to a small group of users to test—it’s unclear when a wider rollout for more chatbot users will happen. The feature is expected to be available for all chatbot users, not just subscribers to ChatGPT Plus.
WIRED received early access to the feature from OpenAI, and I’ve spent the past few days trying it out to better understand this new functionality and to collect some advice you can use to get started. While a few specifics may change before the feature is made available to everyone, here’s what you can likely expect from ChatGPT’s Memory.
Upon the arrival of ChatGPT’s Memory to my paid account, I was met with a pop-up introducing the new feature and providing directions on how it aids in recollecting certain details from conversations. It’s essential to note that Memory function is already switched on, and if you wish to turn it off, go to Settings, select Personalization, and switch off the Memory function by sliding the toggle to the left.
Although there’s slight overlap, ChatGPT’s Memory is not designed to substitue the “custom instructions” feature.
It’s quite easy to add personal details to ChatGPT’s Memory: Simply interact with the bot. As interactions with OpenAI’s software increase, ChatGPT learns more about the user like their name, location, favorite movies, and disliked foods. While the bot gathers this data passively, explicit commands can also be used. Starting a prompt with “Please remember that …” or similar phrasing can add specific information into ChatGPT’s Memory.
Erasing memories from ChatGPT is simple as well. To see everything that the bot recalls, click on your username, navigate to Settings and Personalization, and then hit the Manage button located at the bottom. Then, select the delete icon next to any memory to remove it. Alternatively, Clear ChatGPT’s Memory can be selected to totally reset the bot’s memory.
If you already have a grasp on what the bot remembers about you, you can request that certain things get deleted or adjusted through conversations with the AI. Prompt ChatGPT to change what’s saved in the Memory with commands like, “Forget everything you remember from this chat,” or “Adjust my current city from San Francisco to Sacramento.”
David Nield
Elissaveta M. Brandon
Simon Hill
Parker Hall
What ChatGPT remembers about you can be adjusted in Settings or by chatting with the bot.
My first reaction to the announcement of this new feature was apprehension over OpenAI storing personal info about me and potentially using my personal details to improve future AI models. OpenAI’s FAQ for the feature even reads, “Memories may be used to train our models depending on your choices and plan.” I’m glad to see that users can opt out from training future models—by going into Settings and Data controls—while still having the Memory feature enabled, though this doesn’t assuage all my privacy concerns.
Want to avoid the Memory feature for a few chats but not turn it off altogether? OpenAI is also testing out a temporary chat tool. Enable it by opening the AI model menu, labeled either ChatGPT 3.5 or ChatGPT 4, and select the Temporary chat option. These conversations will still be saved for up to a month by OpenAI, but they won’t be included in model training, the bot’s Memory, or your chat history.
Do you share a paid ChatGPT Plus account with friends or family members? Then it’s probably best to keep the Memory option turned off. With Memory activated, the chatbot might blend all of the details from multiple interactions into one composite understanding of who the user is.
OpenAI will continue to experiment with what ChatGPT is and isn’t allowed to remember, but certain topics are already off limits. For example, you can’t ask the chatbot to remember your Social Security number or passwords. This is also worth highlighting since it may feel counterintuitive: Deleting a conversation from ChatGPT doesn’t also destroy the memories associated with it. So, you have to scrub your memories as well as your chats if you’re looking to clear out the details.
I was decently impressed by the new feature during my initial interactions. One of the major issues with utilizing chatbots for productivity has been their lack of automatic retention. Having to remind ChatGPT about crucial details over and over is a frustrating time suck, and only power users are likely to go in and manually tweak the custom instructions, where you can save personal details and change ChatGPT’s cadence. It feels much more seamless to control what the bot remembers through text prompts.
As the AI race continues, chatbot companies are likely to continue with this personalization trend by offering additional features that adjust the outputs based on what the software knows about you. For now, it’s the user who may improve at AI prompting by participating in multiple conversations with the tool. In the future, the chatbot may return the favor and get better at providing satisfactory, context-rich answers to your questions the longer you use it.
