What type of content do you primarily create?
Remember that feeling when you first discovered ChatGPT? The excitement, the possibilities, the...constant need to re-introduce yourself? Well, OpenAI heard our collective sighs and introduced one of its biggest feature additions: Memory.
As someone who's been eagerly anticipating this feature, I dove headfirst into testing it for the past two months—and came up, honestly, a little dissatisfied. I've found it is one of the most difficult features to use effectively. Its defaults can be both frustrating and counterproductive.
All is not lost, though! With a bit of guidance, you can tweak ChatGPT’s memory feature so it’s genuinely useful. Here are the insights from my two months of testing.
How does ChatGPT's memory work?
The basics
"Without memory, every interaction with AI feels like a first date—starting from scratch, introducing yourself repeatedly, and hoping for a spark. With memory, however, it's like catching up with an old friend who knows your preferences, quirks, and the nuances of your conversations."
At least, that's how OpenAI describes it. I'd say that your "old friend" ChatGPT has a few quirks of its own that you need to remember about it too. It's less "BFF who finishes your sentences" and more "that friend who forgets what your job is, but remembers that random comment you made about bell peppers three months ago." But I digress.
Before we dive into the nitty-gritty, let's cover the basics. Under Settings > Personalization, you can also toggle whether Memory is enabled for new chats. If you have memory enabled, it will automatically add memories when it deems something relevant. More on this later.
On desktop, you can see memory entries being added by hovering your mouse over them as they generate. Memories are accessible under Settings > Personalization > Memory > Manage. Here, you can read individual memories, delete them, or clear all memories.
You can store about 1,200–1,400 words of memories in total. (It doesn't matter if they're in short or longer entries.) Once your memory is full, ChatGPT can't add more memories unless you delete some. This means you need to regularly review and delete irrelevant memories to keep the feature functional.
Managing memory
To have more control over ChatGPT’s memory, work with it in a chat. You can also ask ChatGPT to tell you all the memories it has stored and request the removal of specific references. (Tip: ask for them verbatim, or else ChatGPT will summarize the content.) In the chat, ChatGPT can also edit memories instead of just deleting them.
If your memory is completely full, ChatGPT can't edit it, even to delete something. This feels like a bug so maybe it'll get fixed later. Regardless, to keep the memory manageable, avoid letting it reach full capacity.
Memories are stored separately from chats and are organized chronologically. Although the dates are hidden in the settings menu, you can ask ChatGPT to provide them, which is useful if you want to see how far back the memories go.
A few housekeeping items:
- Deleting a chat doesn’t remove the associated memory; you need to delete the memory explicitly.
- Once deleted, a memory can’t be recovered.
- Temporary chats do not use the memory function.
- The information stored in memory might be shared with OpenAI, as per your data-sharing preferences. You might want to turn this off if privacy is a concern.
How to optimize your ChatGPT memory
What does ChatGPT remember?
The types of things that ChatGPT will add to its memory fall into three main categories:
- Personal details: Facts about you, such as your location, family, profession, and your general interests.
- Preferences: Your likes, dislikes, and even your feedback on ChatGPT's responses.
- Activities and projects: Details about your activities, travel plans, ongoing projects, and their priorities.
I found that if I just let the memory ambiently save and update without deliberately optimizing it, it wasn’t all that useful. It tended to add strangely specific preferences. My personal favorite? When I was tweaking a recipe, it decided to remember that I "Like to use both green and red bell peppers in recipes for their combined verdant flavor and sweetness." Thanks, ChatGPT, but I'm not sure that my preferences regarding bell peppers is crucial information for our future conversations!
Because I was getting frustrated with the tool storing memories I thought were superfluous, I added instructions about what I did and didn't want it to remember to my custom instructions—which it then completely ignored. It seems that you can’t directly control the way it decides what memories are stored, at least for now.
Deliberately add core memories
If you're first starting out, it can be daunting to figure out what to add, so it's tempting to just start using the tool and get it to add things as you work with it, like I did. But I'd warn you against using it like that—for the many reasons I get into below. Instead, start out by giving it some core memories to work from.
Here are the general categories of things I found useful to add:
- Current projects and priorities: What's keeping you up at night (in a good way)?
- Areas of interest: Your passions, hobbies, and fields of expertise.
- Challenges or obstacles: What mountains are you trying to climb?
- Your bio & general information: Basic information about yourself to help ChatGPT frame its interactions with you.
You can also add things like your preferences and feedback for how the tool produces responses, but these are usually best put into your custom instructions.
I found it most useful to add information about my most important projects and my priorities. For instance, I'm currently working on a science fiction story called The Trillionaires, with a goal to make meaningful progress over the summer. As soon as I asked ChatGPT to remember that this project was my priority, it suggested I work on it in nearly every interaction. This was actually pretty useful as it reminded me of my goals and prevented me from going off on other tangents. But beware—if you add a challenge or priority, it will hound you about it forever.
Avoid overloading the memory
While adding useful memories is important, don't overdo it. Memories have a significant impact on the suggestions ChatGPT provides. One of the most frustrating things about working with memories was how it pigeonholed its suggestions to the stored information. For instance, when asking ChatGPT about a music video I was storyboarding, it suggested I incorporate my personal preferences for clothing styles.
That said, I found it sometimes came up with some interesting mashups. I do a lot of fashion design and pattern ideation with ChatGPT so I found it delightful that it suggested we do some designs inspired by the story.
Watch out for AI brain fog
ChatGPT tries to add useful information to its memories, but it doesn't always hit the mark. Sometimes, in an attempt to be concise, ChatGPT will summarize things so much that it loses critical details. For example, when I talked about my projects, it recorded them in some form but missed the finer details.
I have found that when it doesn't have a good memory to go from it can try to guess—so it will sometimes make something up rather than confess that it doesn't know. Here, for instance, recording existing songs or performing live are definitely music-related projects I've never considered, since I'm still at the basic level learning composition.
Keep it succinct and tidy
Given the limited memory space, especially if you're juggling many projects like me, it's crucial to keep your entries succinct and organized.
If you work with it in the chat, ChatGPT can update its memory. But watch out! It's inconsistent in executing your requests. Sometimes it fails to update despite claiming otherwise.
You can also ask it to shorten a memory to make it more concise or consolidate multiple memories (psst, use Chain of Density for this). This is very useful because ChatGPT's default writing style isn't particularly dense, and it can waste a lot of space if you let it save everything it wants.
For instance, I got it to rewrite the lengthy "Is starting a new project called "Humanity's First O'Neill Cylinder which will be a video essay discussing what is needed to create a habitat at L4 or L5" (28 words) to the succinct "New project ‘Humanity's First O'Neill Cylinder’ is a video essay on creating a habitat at L4 or L5" (18 words, a reduction of about one third).
Unfortunately, it's inconsistent even in this task. At times, it deletes memories as instructed, but other times it declines or appears to execute the command but doesn't actually do it. Always double-check to ensure it's adding and deleting correctly.
Don't throw out your TL;DRs
If you're like me and work on numerous projects, you'll almost certainly still need to use your project TL;DRs—that is, a brief explanation of the project you can paste into a prompt. I found that when ChatGPT added project information to the memory, it saved small snippets across many memories. When these were consolidated, they remained lengthy. I found it was easier to keep a TL;DR for key projects and use it as part of the prompt rather than trying to tuck the info away in memories.
Add “pointers”
Memory fills up quickly, so here's a cheat: instead of storing all information in memory, use a link that points to a file or webpage. I have several links stored in memory, including to my personal and corporate websites. This allows ChatGPT to retrieve more information about me and my projects without directly storing everything.
You can also point it to specific documents. However, they need to be accessible on the web for ChatGPT to retrieve them, so be mindful of what you include.
Is memory worth it?
ChatGPT's memory feature is cool but it requires active management and a healthy dose of fact-checking. You can't just "set it and forget it." But, if you set it up right, it does indeed make it a more useful and personalized tool for your needs.
But I gotta go. ChatGPT told me I'm supposed to be working on The Trillionaires.
