Ep 259: ChatGPT’s New Memory Feature – the good and the bad

Resources

Join the discussion: Ask Jordan questions on ChatGPT

Related Episodes: Ep 217: 7 Steps on How To Use ChatGPT in 2024
Ep 246: No that’s not how ChatGPT works. A guide on who to trust around LLMs

Upcoming Episodes: Check out the upcoming Everyday AI Livestream lineup

Connect with Jordan Wilson: LinkedIn Profile

Unveiling ChatGPT’s Memory Feature: Scenario-Based Review

Recent developments in AI technology bring exciting features to the table. One such update is the introduction of a memory feature in ChatGPT by OpenAI. This dynamic addition addresses common user complaints about model forgetfulness, making conversations more engaging and contextually accurate.

Memory Management and Temporary Chats

The memory feature allows users to include specific information for the AI system to remember. The power to manage these memories, whether they view, delete individual items, or clear the entire memory, rests with the user. ChatGPT also addresses any incidents of conflicting memories, deleting incorrect data when the user provides correction.

The application also offers an option for temporary chats. These occur without any memory retention and won’t include the session in chat history. This feature is great for those who want a quick, non-embedded conversation, providing the liberty to operate without the model remembering anything.

The Drawback: Slowed Response Time

One downside observed with this feature is an increase in the chat response time. Testing shows a 30-50% rise in these times, potentially a trade-off for the added advantage of memory retention in interactions.

Visual Clarity: UI/UX Enhancements

Design plays an integral part in machine-user interaction, and the UI/UX of memory feature is noteworthy. Clear indicators differentiate temporary chats from regular ones, ensuring the user is aware of the chat’s nature and setting. However, there is space for improvement, with possibilities in automatic memory commitment from the AI or more customizable memory toggling within individual chats.

Convergence of Memory and Custom Instruction

Though the chat memory feature carries its charm, its functionality could benefit from a more integrated approach. A combination of memory and custom instruction features would allow the model to offer more personalized responses, creating a detailed, rich user experience.

Applications: Ideal for Limited Use Scenarios

During its initial phases, the new memory feature seems best suited for limited use of ChatGPT. Its utility within such scenarios is its ability to carry important aspects of one conversation to the next, lending continuity.

For the future, a promising prospect lies in introducing user profiles for memory, enabling a more segmented and personalized memory based on different chat histories. While the addition of a memory feature brings us one step closer to an AI providing a more human-like interaction, these continuous advancements continue to amaze.

Topics Covered in This Episode

1. ChatGPT’s New Memory Feature
2. Benefits and Drawbacks of the Memory Feature
3. Suggestions for Improving ChatGPT
4. Application and Actual Use

Podcast Transcript

Jordan Wilson [00:00:16]:
One common thing that people complain about, ChatGPT and other large language models, is they say, it forgets things. They say, oh, this this model doesn’t have a good memory. It gets confused. Well, there’s a new chat g p t feature that might help this. It’s called memory, but I think it’s maybe a little more marketing than actual function. So we’re gonna be talking about that today and more on everyday AI. What’s going on y’all? My name is Jordan Wilson, and I’m the host of everyday AI. We do this every single day.

Jordan Wilson [00:00:51]:
It’s a livestream podcast and free daily newsletter, helping people like you and me just learn and leverage generative AI. So there’s always new models, new new new tools, new features. So we’re gonna be talking about that today and see if we can help you, get a little bit more, out of chat gbt in this instance to grow your company and to grow your career. So, if you’re new here, please go to your everydayai.com. Sign up for that free daily newsletter. Maybe you can’t make the the live stream every day. Check us out on the podcast. Maybe you’re on the podcast and you wanna come hang out at the live stream and ask questions.

Jordan Wilson [00:01:21]:
You can do that too. So always make sure to check the show notes in the podcast, for links to do that. So before we get into this new feature from ChatGPT, and, you know, we’re gonna go over some of the pros and the cons. So, you you know, I’m gonna tell you right now, I’m not gonna be using this a whole lot or at least not relying on it because it’s actually going to be, enabled by default once you do enable it. But we’re gonna get to that in a couple of minutes. But before we get to that, we’re gonna start as we always do. Well, not as we always do with the AI news. Actually, a quick reminder.

Jordan Wilson [00:01:54]:
Oh, no. Let’s just go into the AI news. Alright. So, there was a an athletic director in Baltimore was arrested for using AI to fake an audio recording. So a Baltimore high school athletic director was arrested for allegedly using AI to fake an audio recording of the school’s principal, leading to his temporary removal. So the audio clip contained racist comments against black and Jewish individuals causing significant repercussions within the school community. There was a 3 month probe by local police, and the FBI concluded that the audio was forged, by using AI tools by this high school athletic director. So apparently, what happened is the principal had initiated an investigation into the athletic director for potentially mishandling school funds, leading to this retaliatory fake audio creation.

Jordan Wilson [00:02:49]:
You know, kind of a kind of a wild story, but I actually think this is gonna become pretty commonplace with how good, you you know, deep fake audio is and how easy it is now to, to replicate someone’s voice. Alright. Next piece of AI news. US and China are holding their first high level talks on artificial intelligence. So the US secretary of state, Anthony Blinken, announced that the US and China will engage in their first intergovernmental dialogue on artificial intelligence in the coming weeks. So the talks aim to address the risk and safety concerns associated with advanced AI technologies. And despite the discussions, no update was provided on the future of Chinese social media giant TikTok during the recent talks in Beijing. Yeah.

Jordan Wilson [00:03:33]:
That’s that’s gonna be obviously a hot topic here in the next couple of weeks. And the agreement also includes efforts to stabilize and develop US China relations, expand cultural exchanges, and to consult on international and regional issues. Yeah. I think the the the conversation on artificial intelligence is going to be an important one for the 2 different heads of state, to have. You you know, I’d say right now, you you know, the US is in 1st, 2nd, and 3rd place when it comes to, artificial intelligence and generative AI developments. But I’d say China is probably right right there after. Right? It it isn’t really a race, so far, I would say, but China is, I would say, top 3 world power when it does come to artificial intelligence, specifically when it comes to large language models. Speaking of large language models, last but not least, Elon Musk XAI startup is nearing $6,000,000,000 in a new fundraising round.

Jordan Wilson [00:04:26]:
So XAI xAI Corporation, which is Elon Musk AI startup, is on the verge of closing a $6,000,000,000 fundraising round according to reports. The fundraising would value the company at $18,000,000,000 showcasing significant investor interest in the AI technology. So Sequoia Capital, big name there among other investors, are set to participate in this funding round. So this move positions XAI to compete with, you know, more established players like OpenAI, Google Gemini, Anthropics Claude in the AI market. So we’ll see how this shakes out. I I’ve said this, you know, I had a, you know, my show kind of my prediction show in Jan in December for 2024, and I said, you know, hey. There’s gonna be a lot of noise around, XAI and and Grok, which is the XAI chatbot platform. I’m like, you know, probably you shouldn’t pay attention to it.

Jordan Wilson [00:05:18]:
I I still think that that’s just my personal opinion. I don’t see a lot of utility in this specific large language model, considering a lot of it is trained off of Twitter data, which is not the best data. Alright. There’s gonna be more on those stories and a lot a lot other things that we dive into in the newsletter. So if you haven’t already signed up, make sure you go do that. And, hey, did you did you not see this yesterday? I know we had a long, you know, super long podcast for our 1 year anniversary. And, you know, maybe the newsletter went out a little late, but, hey. If you know, like, 2 friends right now, I think the the the the leader, in our giveaway that we just launched was 2.

Jordan Wilson [00:05:56]:
So we are giving away, Meta’s, a AI powered Ray Bans smart glasses, and our giveaway ends Monday. So all you have to do is refer one friend to be entered, but the person who gets the most referrals is going to get their pick from our prize pool, which does include the, you know, Ray Ban glasses. So if you’ve looked at those, you know, like, oh, those are cool. And we talked about on the news here the other day. Now they have multimodal input and output. So or sorry, multimodal input, which is really, really cool. Hey. If you know 3 people, you’re gonna be in the lead.

Jordan Wilson [00:06:29]:
So make sure if, to check our newsletter if you haven’t already. You’re gonna get a unique you for, referral code and you can go in and, hey, If you refer 3 friends as of, like, yesterday, you’d be in the lead. Alright. So let’s get into this new chat g p t feature, and good morning, from everyone joining us live. Doctor Harvey Castro said, hey. He just got the memory feature today. You need to test it out. Alright.

Jordan Wilson [00:06:53]:
Well, hey. Thanks thanks for letting me know, Harvey, because that’s what we’re gonna be doing today. We’re not gonna be testing it live per se, but we’re gonna be going over, what you need to know. Hey. Michael. Michael from YouTube said that he just referred someone hoping he wins. Well, hey, Michael. If you know 1 or 2 more people, you’ll probably be in the lead.

Jordan Wilson [00:07:10]:
So alright. So let’s talk about this new memory feature from ChatGPT. So, I’m gonna tell you what it is, what it isn’t, how it works, and, suggest to you if if you should use it. So here’s how you will know if you have the feature. I wish there was a better way, but essentially, you just get a pop up. So, you know, if this does sound like something that you’re gonna be taking advantage of or if you’re a pretty regular ChatGPT user, make sure to look for that pop up. Also worth noting, if you use ChatGPT on different devices, which is like what I do, you only see the pop up once. Don’t worry.

Jordan Wilson [00:07:46]:
I’m gonna share it here, but this kinda just gives you a good overview of what this feature is. So it does a couple of things. I’m gonna read, kind of the, the marketing speak here. But don’t worry. We’re not just gonna tell you what it is. We’re gonna dive into it. So, ChatGPT says, that it now has memory access across chats. That’s the important thing here.

Jordan Wilson [00:08:08]:
Right? Before, you know, whenever you would start a new chat inside of chat g p t or any large language model, one of the biggest things is it doesn’t remember anything. So let’s say that you’re, you you know, creating a chat, in any model and, you know, you’re you’re you’re getting some good outputs. Well, what happens then if you go into any other chat, It doesn’t know anything that was going on in that chat. Right? So let’s say you’re telling it about your company. You’re working on a research project, etcetera. And then if you start a new chat that’s kind of related to that, and if you wanted to recall some of that information, you can’t. So that has been one of a pretty common downside of large language models is they think of it as siloed information. So, you you know, when you’re clicking new chat or if you go find an old chat and pick up the conversation, all of that information is siloed, which is obviously a huge downside.

Jordan Wilson [00:08:55]:
So this new feature from open AI inside ChatGPT looks to change that. So essentially it says you can keep the conversation going. It says ChatGPT will, carry what it learns between chats allowing it to provide more relevant responses. It says it becomes more helpful over time. So it says chat gbt will, become more helpful as you chat, continually remembering details and preferences. And then it says manage what it remembers. It says you’re in control. Review and delete specific memories, restart from scratch, or turn off the feature in settings.

Jordan Wilson [00:09:28]:
Alright. So, yeah, I’m I’m curious for for the rest of you joining live if you have this feature yet or if, what questions that you have about it. Alright. So we’re gonna get into, kind of the details of how it works. We’re gonna give you a guide, but if you do have questions for our livestream audience, make sure you get them in now. But if you’re listening on the podcast, we always leave a link back to this kind of LinkedIn thread. So, if you are maybe driving or walking your dog and I say something, you’re like, wait. That doesn’t make sense.

Jordan Wilson [00:09:54]:
You can always come back and ask me, and I’ll I’ll answer. Alright. So here’s what you need to do. Well, number 1, you need to enable it in the settings. Alright. So, this is actually in a different spot than when they first previewed it. And this is also kind of important to to talk about. So this is one of the first features that OpenAI has posted publicly about before they did a general rollout.

Jordan Wilson [00:10:16]:
So they actually posted about this 2 months ago, and a lot of people forgot about it. You you know, OpenAI put out, some information, I think, on their on their Twitter accounts, on the blog on on on their blog posts, and this was 2 months ago. So normally, when OpenAI posts about something, they kind of just release it within a day or 2, or it has already been a release and it’s just a gradual rollout. So, if you do have a paid chat gbt account, I do believe it is only for paid users. They were testing it with free and paid. So if you do have a free account, you go in and if you see this memory feature, let me know. But I do believe it is only rolling out to paid users. So like I said, when you log in, maybe today or tomorrow, the next time you go in there, you’re gonna get this pop up, and then you can go into settings.

Jordan Wilson [00:11:00]:
So previously, this was its own dedicated, kind of side panel. So you didn’t even need to go into settings first. So they didn’t move it. So you do need you need to go into your regular settings, and there will be a personalization, tab. Okay? And you’ll see you can still have custom instructions going. So we’re gonna go over kind of the differences and the pros and the cons of working with each or working with 1 or the other. But then you’ll see there’s a simple memory toggle. So if you do maybe miss it on that pop up, on that pop up and you don’t enable it, or if you’re starting to use it and later you’re like wait this stinks, don’t worry you can always toggle it off.

Jordan Wilson [00:11:35]:
Alright. So that’s step 1. You need to enable it in the settings. So here’s what it’s going to look like in action. Alright. So I just did an example. I just started playing with this yesterday. We obviously had our big, you know, 1 year, anniversary show, yesterday.

Jordan Wilson [00:11:51]:
So I’ve been a little busy. You know, we had the show of, will AI take your jobs? It was our longest show ever. So make sure you go check that out if you haven’t already. But so when you go into ChatJPG and you enter something, essentially, you’re gonna get this new little prompt, And it says memory updated. So in this example, I didn’t have anything in, you know, kind of the memory bank, so to speak. So, you know, think of it like this. You bank memories. So if we literally think of the the silo, you know, the silo analogy that normally they work in silos.

Jordan Wilson [00:12:23]:
So think of it like this. You’re gonna have a base that’s below all of your silos. So it doesn’t matter, and this memory bank is that base. So it is going to, selectively pull things out, of your chats, and then you can go in and, manage those memories as well. But this is essentially kinda like what it looks like for our livestream audience, is it just committed these things to memory. So I just said in a new chat, I said, my name is Jordan. I like deep dish pizza. I’m from Chicago.

Jordan Wilson [00:12:52]:
And I said, people call me a robot sometimes. So right after I entered that, a new pop up came. So I’m glad that ChatGPT did this piece. Because when this first came out, I did a little video review on it, at least on the blog post because they shared screenshots, they shared video recordings, and I said, this is terrible. I said, if OpenAI actually releases it like this, this is gonna make ChatGPT bad for everyone. Because in the original release, there was no pop up kind of notification in your chat. So you could be chatting with ChatGPT, put in whatever information, and it would just commit things to memory. And at least in their, initial release, that very few beta testers had access to, you didn’t know when things were kind of, put in the memory bank versus when they worked, which can be problematic, and we’re gonna talk about that.

Jordan Wilson [00:13:40]:
So I’m glad they made that update. You know, maybe me bugging, OpenAI so much, about certain things. Who knows? Maybe, maybe that influenced them. But, now at least in the chat, it will tell you, memory updated. And then you get this little new pop up that, you know, you probably haven’t seen before, and it kind of shows you what was committed to memory, and then it gives you an option right within the chat to click on manage memory, which is pretty nice. So as an example, I said, my name is Jordan. I like deep dish pizza. I’m from Chicago.

Jordan Wilson [00:14:08]:
People call me a robot. So there’s a little, UI UX tweak here, that I think that they should make because I gave it multiple facts. And at least in this response back from ChatcheBT, it looks like it just committed it as one memory because it says Jordan likes deep dish pizza, is from Chicago, and is nicknamed robot. Even though I didn’t say I was nicknamed robot, I just said people call me a robot. So, little, in the details there, but, still. Alright. So now I can go in and manage that memory if I want to. So here’s the other thing that you need to know.

Jordan Wilson [00:14:44]:
So now I opened a brand new chat. So we just committed that to memory. I’m gonna show you how you can manage or talk to you and tell you how you can manage that memory in the past, but I wanted to skip to the end and say, hey. Here’s how that works. Right? So, I kind of put put those things into ChatGPT. ChatGPT decided this is an important detail that I’m going to commit to memory. And, again, it will do that automatically. You can tell it.

Jordan Wilson [00:15:06]:
You can say, hey. You know, put that in my memory, and it will do that. But, you know, especially early on, it’s gonna be grabbing a lot of different things, in your early on chats once you first enable of this and just committing it to your memory. But then what happens? You know, let’s skip to the end. So I just open a new chat, and I said write a poem about my favorite food. So in this case, I didn’t have to mention a memory. I didn’t have to tell Chad GPT what my favorite food was, and then it responds with, here’s a poem about deep dish pizza inspired by your favorites. Right? And then it even talks about their, about Chicago.

Jordan Wilson [00:15:40]:
Right? Because I said I’m from Chicago. So you can see right away how this could be useful. Right? I’m sure that so many of you, have to spend a lot of time in ChatGPT or any other large language models. Right? Kind of re reminding a chat. Like, hey. Like, I’m a I’m a I’m a researcher. This is what I do. You you know, this is the cut this is the type of information that I care about.

Jordan Wilson [00:16:03]:
This is how I want your outputs to look. So all of these different things can be committed to memory. Alright? And then you’ll see, you know, it’s pretty useful in that regard. Right? That you don’t have to, you know, always remind it of certain things about you. Alright. So then you can manage your memory from the settings. So, in the same way that you enabled it, you can also go click the manage button and then you can manage it. So, I was I have some other kind of memory, things here.

Jordan Wilson [00:16:30]:
I was trying to trip chat gbt up. You know, we always try to test these things and and break them. So as an example, you know, here and in another chat, I said, hey. Bob hates pizza. So I was seeing if it could, if it was gonna get mixed up. And if I said, hey. Write it, you know, write a poem about my favorite, favorite food. If I gave it kind of conflicting, you know, points of points of information.

Jordan Wilson [00:16:52]:
Right? I said, hey. My name’s Jordan. I love pizza. And then I just said a imaginary story about this guy named Bob who hates pizza. And I was seeing if it would trip up, but it didn’t. But, anyway, so you can go in and manage your memory. And you’ll see, luckily here so I said when it originally committed that kind of statement to memory, it looked like it was putting it in as one big chunk, but it didn’t. So it made separate kind of, memory instances.

Jordan Wilson [00:17:14]:
So it said Jordan likes deep dish pizza. That’s one memory. Jordan’s from Chicago. That’s another memory. Jordan is nicknamed robot. That’s another memory. So what you can do from this manage memory section is you can delete them all, individually or you can clear, the memory. So one thing that I would like to see is the, the ability to edit a memory.

Jordan Wilson [00:17:37]:
So, like, even as small example there, when it says Jordan is nicknamed robot, I’m not. I was just saying people call me a robot sometimes. So I’d like to be able to edit that and update it. Or even from this instance, I think it would be great if you could add new memories here, but you cannot. So that’s something important to keep in mind. Once chat g p d has kind of committed something to your memory bank, you can’t edit it. Also, you can’t go into manage your memory and add or, you know, can’t edit things, but you can’t add things either. All you can do is delete them.

Jordan Wilson [00:18:06]:
So, you might have to. If you are trying to commit certain things to your ChatGPT memory, you can obviously tell it and say, hey. Commit these things to memory. List them in bullet points on new lines, and it will kind of put it in that way. So that’s probably how I would recommend, people use this versus just giving, you know, ChatGPT the freedom early on. So if you’re gonna go out and, you know, test this today or if you wanna use it early next week for a project, I would suggest doing it like that. Saying, hey, ChatGPT, commit these things to my memory. Bullet point them.

Jordan Wilson [00:18:37]:
Make them very specific. Otherwise, what’s gonna happen is ChatGPT is gonna just gonna take some, it’s gonna take some, creative freedom in thinking, what you want committed to memory and even how it interprets certain things that you say. So that’s important. You know, the the the managed memory part is a little tricky. So, great couple great questions here, from from Wuzy and Harold. So thanks for the questions. So Wuzy said interesting. In your opinion, as you’re using it, does it feel like it’s trying to commit the appropriate things to memory? So I’ll say yes.

Jordan Wilson [00:19:12]:
You you know, again, but early on, you know, I don’t have a lot in the memory. I just started playing with this, you know, probably for, an hour so far because, yes, it did just release to most people within the last day or so. So I haven’t had a ton of chance, woozy to, to, play with it, but I do have some thoughts on that here later, in my little presentation. And then Harold says, I’m guessing if you correct it, that it’ll override its training information. Yes. That’s a great question. So it will, delete those pieces of memory. So let’s say it gets something incorrect.

Jordan Wilson [00:19:44]:
So even if you are in the chat and say, you you know, I can say, oh, I’m not nicknamed robot. People just call me that. You know, update the memory. It will actually delete the old one, so there won’t be contradictory things. So, yes, ChatGPT will, on its own kind of manage the memory on its own. So it’s not like, you know, if I say, oh, people call me robot. It got it wrong. It said Jordan’s nickname robot.

Jordan Wilson [00:20:06]:
So it’s not gonna have 2 of those different things. It will automatically delete the one, you know, if I tell it to or if it realizes that it made a mistake in committing something to memory. Great questions. Great questions. Alright. So let’s get this thing, let’s keep going because, here’s some important things. So, kind of number 5 thing that you need to know is you can start a temporary chat without memory. So you are gonna have a new option when you go into ChatGPT, after you’ve enabled memory is there’s something called temporary chat.

Jordan Wilson [00:20:37]:
So this is important, and this is kind of where custom instructions actually come into play, and I’m gonna tell you about that here in a second. So if you choose temporary chat, here’s some of the toggles that come up. Also, I apparently like, I have 4 chat accounts. Apparently, I need more because I have 2 on Teams and I have 2 on free, but I just realized I don’t have any on the normal chat gbt plus. So I don’t know if this current feature so, hey, if you do have normal chat gbt plus, which is $20 a month, I should actually explain them. So our team has ChatGPT teams. So it’s a little different. You pay an extra $10 a month, and there’s actually a lot of nice features.

Jordan Wilson [00:21:16]:
There’s, some better sharing features, some more data control. And the thing that I actually like is you get double the limits. So you don’t have that 40, message limit within a 3 hour period, which is what the ChatGPT plus account limits are. So if you’re on a Teams account, you actually have double that. You have 80 messages within a 3 hour window. But what I just realized is I believe this feature here may only be available, to those on ChatGPT teams. So apologies, but, hey, if someone could let me know. So the temporary chat has something at the very bottom.

Jordan Wilson [00:21:49]:
But let me just tell you what it is. So it says not in history. It says temporary chats won’t appear in your history. For safety purposes, we may keep a copy of your chat for up to 30 days. On a temporary chat, there’s obviously no memory. So it says ChatGPT won’t use or create memories in temporary chats. If you have custom instructions, they’ll still be followed. Then it says no model training.

Jordan Wilson [00:22:11]:
So temporary chats, won’t be able, won’t be used to improve our models. Another important thing kinda related to this. So, GPTs still work normally. However, you cannot commit things to memory right now while chatting with a GPT, just because for this very issue. Because, with the temporary kind of, chat and what that, kind of gives you a little, additional you you know, think of it like an incognito browser almost. So you’re saying, hey. When I’m, you you know, doing all this chatting, it’s not gonna commit things to memory. So right now, at least as of my testing yesterday, the memory feature, if you are quote unquote talking to a GPT, will not work mainly because, you know, the 3rd party GPTs, it sounds like there’s probably some some data and, security, things that just aren’t in place yet, to be able to still have that, you know, quote unquote temporary chat functionality, and still chat with third party, GPTs.

Jordan Wilson [00:23:08]:
So that’s important to keep in mind. So let’s get back to the temporary chat and see what that looks like. So I actually like the, the the the UI of this. Right? So when you have a temporary chat, I like this. It tells you at the top, so that’s good. So you don’t have to say, okay, is this a temporary chat where I’m not using the memory? But so in this case, I said, what is my favorite food? And then chat g p t says, I don’t know what your favorite food is, but if you tell me, we can chat about it. So the good thing I like as well is at the bottom, the kind of text input box becomes this kind of a dark gray or black color. So So it’s nice.

Jordan Wilson [00:23:41]:
So you have multiple kind of reminders that this is a temporary chat, and you’re not using that memory feature. So you might ask why. Right? Well, there might be times when I don’t want chat gbt to be maybe knowing about I’m writing about pizza or I’m from Chicago. Right? So let’s say I’m you you know, or maybe you are writing something for a client. Maybe you’re using this to write for, you know, multiple clients. So that’s something with, this this memory feature and, even custom instructions, which we’re gonna get to in a second. That’s something I don’t necessarily like, that it’s it’s not always super easy to toggle this on and off. I wish that there was a an ability to toggle this memory feature on and off within a chat, but you do almost have to go into this own dedicated mode of a temporary chat.

Jordan Wilson [00:24:29]:
Because here’s here’s what I think why that’s problematic. I think a lot of people are gonna start a normal chat. They’re gonna keep going, and then all of a sudden they’re gonna get this crazy response. Right? And they’re gonna be like, oh, you know, I’m writing this for, you know, I’m creating an an SOP for a banking client, and now all of a sudden it’s bringing in my personal information. Like, what’s going on here? And then you remember, oh, I’m in a normal chat, so it’s bringing in all my memory. So you kinda have to start over at that point and go in and create a new chat and make sure it’s temporary. So, I I wish that there was a way to toggle this on and off within individual chats, which is another reason why I don’t use custom instructions, kind of the same same reasoning. So we know how hard it is to learn AI and how to actually leverage it, but that’s what our newsletter is for.

Jordan Wilson [00:26:09]:
So speaking of that, that is one of our last points that we wanna make here in today’s show, is it can be used with or instead of custom instructions. So when you do go in your settings and you go to personalization to look at your memory, you’ll see that you can still use custom instructions and memory at the exact same time. So here’s, kind of the the the difference. Okay? Custom instructions, if you haven’t used them, or if you maybe don’t know what they are, It is literally that. So you are giving instructions to ChatTpT, specifically on how you want it to act or how it should perform.

Jordan Wilson [00:27:07]:
So you’re kind of giving it instructions, and you’re saying, hey. You know, when you respond, always respond in an informal tone. So this is where I think there’s I wouldn’t say confusion, but there’s actually a lot of crossover here. Because when you go into your custom instructions, within ChatGPT, there’s actually, 2 different things. So custom instructions ask you, what would you like ChatGPT to know about you to provide better responses, which is kind of what this whole memory thing is. And then there’s another feature, or another segment, in custom responses that says, how would you like ChatGPT to respond? So with memory, you should think of memory as certain details about whether you, your client, your business, the type of work that you’re using chatGPT-4. So there is still a lot of crossover, and I think it’s actually gonna be problematic for people that like and use custom instructions to use memory because there might be some conflicting or slightly different, custom instructions versus what’s in your memory. So at least for me personally, I don’t like using custom instructions, because again, you have to either enable it or disable it account wide.

Jordan Wilson [00:28:22]:
So at least with the the new memory feature, you can still use memory for some of those things that you might normally use custom instructions for. As an example, you know, all of your replies should be in informal tone. Or, you know, when summarizing a long article, you know, summarize in short bullet points. So you can still add those things to your memory even though that might be something that you would generally use custom instructions for. So the benefit here and kind of one of the reasons why I might be using, this feature when I don’t really use custom instructions is for that. So I can have now different chats that are saved in my account. Some of them are using memory and some of them aren’t. Where at least with custom instructions, you do not have that ability.

Jordan Wilson [00:29:05]:
You have to essentially toggle it on and off for your entire account. So let’s say if you did wanna toggle custom instructions on and you were in an old chat, well, that one doesn’t have custom instructions. So, I I I kind of wish that chat gbt would just merge these things. That would be a lot better. I think you kind of have these 2 different layers of customization. 1, you can toggle with memory. You can toggle it on and off. Well, you can’t really toggle it on and off, but you can choose a temporary chat, which essentially toggles it off.

Jordan Wilson [00:29:39]:
And then you have custom instructions, which again can only be toggled on and off system wide. So, a little confusing there, I’d say, from ChatGPT. I would hope, that OpenAI just combines these two features, But, you know, who knows? So that’s something important. If you do want to use this new memory feature, again, think of it as you’re pulling little bits and pieces of information, whether it’s about you, your company, how you like to receive things. Maybe, you you like similes and metaphors. Right? So you can add a you you know, you can do this in custom instructions, but you can also do it in memory. You can say, you know, when explaining complex topics, use similes or metaphors. Right? So these are things, but they’re going to follow you around, in all of your chats unless you start a temporary chat.

Jordan Wilson [00:30:23]:
And then the last thing that’s important to know, well, at least right now, this new memory feature is really slowing down responses. Again, I’m not sure how it’s working under the hood, but in my testing, it is taking a little bit longer. Even if I’m, you know, asking something that maybe is not related to memory or it’s not pulling the memory, I’m I’m finding the responses a little slower. The reason being now is, you know, you could think of this as, like, super mini rag. Right? You know, you’re putting these, you know, retrieval augmented generation. Right? You’re not putting in, you know, gigabytes of of files, but each time that you’re now interacting with ChetGPT, it is kind of checking with this third party, quote unquote, but it’s checking with your memory bank each and every time. And, at least in my early testing, I did notice that it was markedly slower, not a lot, but maybe I would say response times were taking 30 to 50% longer. It could be because this is a new feature.

Jordan Wilson [00:31:24]:
It’s not fully optimized. It could be there’s an influx of new users trying out this feature. But at least right now, this was slowing down responses. So whether it was committing things to memory, checking if something that I talked about was in memory or not, or replying to something that was in the memory. This new feature was slowing Checkatrade down just a little bit, so that’s something to keep into consideration. Alright. So that is it. I think there’s a couple of questions.

Jordan Wilson [00:31:50]:
Let’s go ahead and get to them. So, Tanya saying doesn’t mean that, doesn’t this mean that ChattGPT could conduct bios, bio sorry. This is very, small small font. Essentially, bio research on us. Yeah. I mean, technically, yes. But this has already been going on anyways. So, yeah.

Jordan Wilson [00:32:10]:
OpenAI can use all of this information for training data. So if you do start a temporary chat, it doesn’t do that. Right? So it’s not gonna be training, you know, its models on your memory and, you know, how it responds to those memories. So, yeah, I definitely and Tanya, thanks thanks for that. I always say never put any, you know, confidential, proprietary, you know, personal health information into, any large language model. But the same thing, you probably shouldn’t be putting in, you know, very personal details because in theory, those would be used to train future models. So that doesn’t mean it’s it’s going to be in its data set, but it is going to be used to train their future models. Jonathan, good question.

Jordan Wilson [00:32:51]:
Says, is it clear how to define when you are using a temporary chat versus a non temporary chat? Thanks for that reminder. Yes. So it does have the little banner up top as well as the kind of, darker, response at the bottom. Whereas, normally, if you’re in a normal chat, it’s kind of the white message input box. So you have kind of 2 different reminders there. Alright. I think we made it to the end. So I hope this was helpful getting this overview of the brand new memory chat.

Jordan Wilson [00:33:20]:
I’ll tell you this. I’m gonna be using this at least early on. I haven’t used custom instructions because I use ChatGPT for a variety of tasks. Right? I’m in ChatGPT essentially all day. And this is where, at least for custom instructions, you know, having this blanket application, to to everything, to all of your chats didn’t really work out for me. So I’ll probably start using this, because there are certain things that I would like to kind of commit to memory, and then having the ability to do this temporary chat, is great as well. So if I don’t want all of those kind of personal memory things to influence, my other outputs, you have that option as well. So I would recommend that most people at least give this a try.

Jordan Wilson [00:34:01]:
It’s gonna work best. So this new memory feature is gonna work best if you have a limited or a more limited, use of ChatGPT. So let’s just say you use it just for your business or just for personal or maybe if you, have clients, you use it just for, helping create content or research for 1 client. But if you’re using it for many different clients or many different reasons or, you know, personal business and side projects and clients. That’s where I think you might run into some issues, and that’s kind of how I use ChatGPT. So I’m gonna give this a try. If you use ChatGPT a ton for many different use cases, I think it’s gonna be a little tricky to manage that quote unquote memory bank. I think it would be nice if in the future that if it gave you kind of profiles to work on and you could choose a profile and, hey, here’s these 20 pieces of memory that go with my, you know, personal profile.

Jordan Wilson [00:34:51]:
Here’s these 20 pieces of memory that go with, you know, the everyday AI business profile. But for for me, I personally use, ChatGPT and other ledge and other large language models for so many varied use cases that I don’t think I’m gonna be able to get a ton out of the memory, but I’m gonna try it out and I’m sure we’ll be, you know, checking in. So that’s it. I’d love to hear your thoughts and what you think if this is gonna be a feature that you use. So if this was helpful, please consider sharing this with your network, leaving us a, review on Spotify or Apple if you’re listening on the podcast. And, hey, do not forget. Do you literally have 3 friends? Do you have 3 people in your network? So literally to enter our giveaway, all you have to do is check our newsletter either from yesterday or from today. There’s gonna be a unique, referral link.

Jordan Wilson [00:35:37]:
Grab that. You can post it on social media. You can email it to 10 people, to 2 people. You can text it to someone. Alright? So right now, all you have to do is refer one friend to enter our giveaway. We’re We’re gonna have multiple prizes, but our kind of marquee prize is these new AI powered meta Ray Bans smart glasses. They’re really cool. They just added this multimodal, functionality so you can put them on and you can say, hey, Meta.

Jordan Wilson [00:36:00]:
What am I looking at? And it’ll tell you. Really cool. And like I said, we didn’t have a lot of entries yesterday, which was surprising. So but that’s good for you. So, if if if you find value in everyday AI, please consider signing up for that, referral contest. Check out the newsletter from yesterday or today. Tell a couple friends about it because, hey. As of yesterday, if you refer 3 people, you’re in the lead for the Meta Ray Ban glasses.

Jordan Wilson [00:36:24]:
So that’s not hard. We all know 3 people. Alright. So, that’s it. Make sure to go to your everyday ai.com. Sign up for the free daily newsletter. Check out our giveaway, which ends Monday, and we hope to see you back for more everyday AI. Thanks y’all.