Memory for ChatGPT: Shaping the Future of Personalized AI

OpenAI has released ChatGPT with memory for everyone, which means you will no longer need to tell it for the 100th time your writing style, company name, etc. This update finally brings the most requested feature for ChatGPT and Generative AI products to all users. Having past chats and user-specified data available for the model to access has taken us one step closer to the “personalized AI” of the future.

• What is this new memory feature and is it the tool that users have been asking for?
• Have other products already created similar products?
• Which product should users explore?

These questions and more will be answered in this article about the state of long term memory in the Generative AI ecosystem.

Memory for ChatGPT

The ChatGPT memory feature can “remember” content from the conversations you have with the model. OpenAI’s blog post gives the example of a user explaining that they prefer their meeting notes to be in a bullet pointed list. ChatGPT will then automatically remember this information and create future meeting notes accordingly. This method is similar to how custom instructions work (which haven’t been talked about much by OpenAI since its release). Unlike with custom instructions where you need to create manual instruction for the model, a user can accumulate these changes naturally through interaction with the model. Since its release, it seems that Memory has surpassed Custom Instructions by being a more user friendly experience while still allowing users the ultimate say in what is being stored. It will be interesting to see if this replaces custom instructions entirely or if there is still a place for them in a world with ChatGPT memory. Another example explained in the blog showcases that ChatGPT will also naturally learn your style and tone preferences and adjust accordingly. Now instead of having to specifically explain and articulate how you want the model to respond, ChatGPT will slowly learn what you prefer, presumably by keeping track of whether you ask it to re-generate something or give the generated answer a bad rating.

How can I access this feature?

Recently, all users were given access to this feature. First select your profile in the top right corner, then click “Settings” followed by “Personalization”.

Finally, select “Manage” and you will see the below screen. This is the menu to control your memory in ChatGPT.

From here, you can view and remove unwanted memories. Currently, you can’t add to your memory through this menu. You will need to manually ask ChatGPT to save a prompt if you’d like it to remember for future reference.

“GPTs will have their own distinct memory. Builders will have the option to enable memory for their GPTs. Like your chats, memories are not shared with builders. “ — OpenAI Blog

Keymate Memory

Now that we know about ChatGPT’s latest updates, let’s look at the newly improved Keymate.AI GPT. While the GPT has multiple core features, we’ll focus solely on the long term memory that’s called “Keymate Memory”. Throughout its development over the past several months, Keymate memory has become a multi-purpose long term knowledge base. Users can save everything from prompts and responses to 50+ page PDFs. The evolution of Keymate memory has resulted in a tool that allows users to choose when they recall saved information versus relying on the models’ training data or browsing the web. The Keymate Memory Manager (seen below) was recently updated and allows users to view and interact with their data in a more familiar way. Users are able to search their long term knowledge base without needing to use the model.

Keymate memory has some similarities to ChatGPT’s new memory feature, but they differ in a few important ways. Keymate is focused on recalling large amounts of specific information while OpenAI’s memory seems more like an upgrade to custom instructions. Both products save and allow the user to access pieces of previous chats. They also both allow you to view and manage your memory (with Keymate having a more visual approach to this as shown above). However, Keymate enables users to save much larger pieces of data like an entire web page, PDF, research paper, Youtube video, or Google Doc. People are using long term memory for things like:

• Finding and saving documents related to research.
• A real estate database that is filled with housing information which the model can access and refer to.
• Web pages that are related to an upcoming article.
• A Dungeons and Dragons character guidebook that’s filled with the details of both players and characters.
• And much more!

The commands to use Keymate memory through the GPT are simple: “add [important text here] to my keymate memory​” or “upsert the last answer to my keymate memory”. To recall saved information, users simply need to make the request in a prompt like this: “find [item_name] in my keymate memory”.

With those simple commands, users can populate their memory with handpicked and personalized data. To help keep your Memory Manager organized, users can create collections and notes that work similarly to folders. Users can also upload new items to their memory directly in the Memory Manager. These actions allow for greater control over what you store. This kind of variety is what makes Keymate Memory so powerful.

You can reference this example chat to see a basic example of how this process works.

A guide on how to manage your Keymate memory can be found below.

So which should I use?

Both! These two features approach the same problem in two very different ways. Neither is better, just different. If you’re searching for information that you want to save and have the model recall later, Keymate Memory is the right tool. If you’d like to increase the effectiveness of your custom instructions to receive more tailored responses, ChatGPT’s memory can help. Overall, the prospect of having both memory types working together is an exciting idea and is an important step towards the future of truly personalized AI systems.