OpenAI upgrades ChatGPT with persistent memory

This is a distinct difference from the way ChatGPT has operated to date, in which the app only remembers information the user provides to it in a single chat session. Once you close that chat session and start a new one, ChatGPT has, until now, started over with a tabula rasa (blank slate) knowledge, other than any “custom instructions” the user may have set up for it in the settings.

Now, with its new memory feature, ChatGPT can help save users time by avoiding them having to repeat information across chat sessions.

However, it is only rolling out to a select group of free and Plus users initially to see how impactful it is. Teams and Enterprise plan subscribers will get the capability at a later stage.

The move to enhance ChatGPT this way comes as competition in the AI assistant space continues to grow intense.

Just last week, Google made an effort to push its AI chatbot front and center with a major rebrand and the inclusion of its most capable Gemini Advanced model under a new paid plan.

What to expect from memory in ChatGPT?

In a blog post published today, OpenAI notes that memory-enhanced ChatGPT will begin picking up details as users chat with it.

The memory will build over time through the natural use of the chatbot. However, if one wants the assistant to remember something specific, they can explicitly ask it to “remember this” and provide whatever information they’d like it to store, as well. 

All the remembered information is stored on the users’ accounts and used to enhance future conversations.

This can prove helpful in various work and personal use cases. For instance, one could ask ChatGPT to remember how they prefer their articles to be formatted and the bot uses the same format without being told explicitly time and again. In other cases, it can remember preferred programming languages to write code in or take note of the charts one prefers to visualize data.

OpenAI says that memories collected from free and Plus users may be used for training its models – just like any other content.

However, if a user wants to avoid that, they can switch to a temporary chat that does not appear in history, won’t use or create a memory and won’t be used to train the models. 

While memory will be enabled by default, users will have the option to turn it off altogether by heading over to the Personalization section in the settings of ChatGPT. There’s also the option to manage memories, which allows users to delete specific memories from their accounts. This can also be done in the natural flow of the conversation by simply telling the assistant to forget whatever the information is.

Focus on safety, broader rollout soon

The move to enhance ChatGPT with memory functions will certainly help users with their work and personal projects. However, implementing it successfully means taking an unbiased approach to decide what should be remembered and what should not. 

OpenAI says it is taking steps to not only assess and mitigate biases but also steer ChatGPT away from proactively remembering sensitive information, like health details — unless explicitly asked to do so.

The company is testing the new capability with free and Plus users of ChatGPT, but a broader rollout covering Enterprise and Teams users is also expected soon. Memories for the latter will not be used for training, OpenAI has noted.

As part of the wider rollout, businesses building GPTs will also get the option to embed memory in their offerings. This will enable the builders to provide more personalized experiences to their customers based on their interactions. However, each GPT will have its own distinct memory from conversations and will tie to what ChatGPT knows.

“The Books GPT helps you find your next read. With memory enabled, it remembers your preferences, such as favorite genres or top books, and tailors recommendations accordingly, without needing repeated inputs,” the company explained.