ChatGPT and Claude's "Memory" Isn't What You Think

Here's something that'll mess with your head: ChatGPT and Claude have zero memory. Every message you send, the model reads your entire conversation from scratch, answers, then forgets everything. When these tools claim they "remember" you, that's not the AI. It's separate software making judgment calls about you, and you never see it happen.

If you've ever wondered why AI gives inconsistent outputs or seems to lose the plot, this episode connects the dots. Alane Boyd and Micah Johnson break down why relying on automatic AI memory quietly works against your team.

In this episode:

• Why AI memory is really judgment calls stacked on judgment calls
• How bad memories become invisible and skew everything after them
• Why context bleeds between clients without warning
• How every team member ends up with a different "source of truth"
• The file-based system that puts you in control of what AI remembers
• Why Biggest Goal shut off automatic memory at the organization level (and nobody noticed)

Memory is just context. Once you control the context, the outputs finally get consistent.

Upcoming Education Sessions:

• Free Claude Cowork Webinar
• Free Claude Cowork Masterclass
• Claude Deep Dive Series

Tools/Platforms Mentioned

• ChatGPT - AI chat interface with automatic memory feature
• Claude - Anthropic's AI assistant with memory functionality
• Claude Desktop - Desktop application for Claude
• Claude Cowork - Business/team version of Claude
• OpenAI Codex - Code-understanding AI model
• Obsidian - Knowledge management and note-taking app
• Google Drive - Cloud storage service for file syncing
• Dropbox - Cloud storage service for file syncing

Learn More about our AI Leadership Workshop

Winning businesses aren’t just working harder, they understand how to use AI strategically. Sign up for the AI Leadership Workshop, a 3-hour expert-led mastermind designed to give leadership teams clarity, alignment, and a practical path to stay competitive.

Book a call to customize your workshop today!

This episode is brought to you by Biggest Goal.

Every quarter your team spends evaluating AI is a quarter your competitors spend shipping. Most leaders feel the pressure but get stuck between ignoring AI and getting it wrong. More tools and more demos won't fix it. What actually works is hands-on training for the people doing the work.

Our workshops align your leadership team on a shared AI roadmap. Our 4-week cohort get your teams building real AI agents, and automation workflows. And our ongoing groups keep leaders and builders sharp long after the cohort wraps.

Upcoming Sessions:

• Claude Cowork Masterclass
• Claude Deep Dive Series

Connect with Us:

• TikTok
• LinkedIn
• YouTube
• Community

Disclosure: Some links are affiliate links, meaning we may earn a commission at no extra cost to you. Thanks for supporting the podcast!

For more information, visit our website at biggestgoal.ai.

Transcript

Micah (00:01)
Here's something that'll mess with your head. ChatGPT and Claude have no memory. None. Every single message you send, the model reads the whole conversation from scratch, answers, and then forgets all of it. So when a tool says it quote unquote remembers you, that's not the AI remembering. That's a completely separate piece of software making a judgment call about you every single time. And you almost never see it happen.

Alane Boyd (00:29)
You know, Micah, it really irritates me because this is such a hard thing to change our mindset about because Claude and ChatGPT lie to us and they call it a memory. So we think that there's actually a memory.

Micah (00:44)
I think it's one of the most misunderstood and honestly confusing features. I hesitate to call it a feature. It is a feature. It can work really well. But this memory thing, right? It's like the core of it is hey, I'm AI. I'm gonna make a judgment call about what I should remember. I'll do it for you. Don't worry about it. But

when you actually look at all this, you go, Well, actually AI is kind of crappy at judgment. So what's going on here?

Alane Boyd (01:15)
It was hard for me to understand at first too, because it calls it a memory when you're in So I think that it's really learning me, and then you always hear people say, well, my ChatGPT knows me or my Claude really knows me. And I think in small ways it does. Like it knows that keep it brief for me. It knows that I like being punchy and straight to the point. But it it really cannot get past like a superficial level, I notice.

Micah (01:40)
Yeah, absolutely. And I think memory is such a misnomer because what that paints the visual picture for humans is that I'm gonna chat with ChatGPT. It has a memory of what we talked about. Therefore, it's going to be able to remember it like a human memory works. And that's just the assumption that a lot of people make, but that is not the case whatsoever.

Alane Boyd (02:04)
So I wanna go through like the downsides of relying on this fake memory. And then we actually have a solution to it, but I want to go through these downsides so people go, oh my gosh, I don't actually want to be just relying on this fake memory. And then we give them the winner of like do it this way instead.

Micah (02:10)
Yeah.

Love it. Love it. Should we just give a brief explanation about how the memory actually works so that people understand? Okay. All right. This was this was I was born for this. This was my favorite. No, I'm just kidding. All right. So the way that the memory actually works is you chat with Claude, you chat with ChatGPT, and at some point it's going to make a decision. It's going to make a judgment call to go, hmm, this sounds important. I think I'll store this away for future access, which at

Alane Boyd (02:29)
that's probably a good idea. Sure. You start there, Micah.

Okay.

Micah (02:53)
At first glance, that seems like hey, that's a pretty good memory. The problem is if you go and chat with ChatGPT and you tell it a joke, or you start talking about race cars, or you want to talk about vacation planning, it's going to make judgment calls on all of those things of what it should store away in its quote-unquote memory. And then you actually want to get some work done. So you ask it about work, and now all the information that it's giving you is colored with

trips to Singapore or Japan and race cars because it saved that in its memory and it thinks that's you. And in some cases, when you're using it personally, that can work really well. but what's happening is it has to make a second judgment call to go, well, now we're chatting about stuff.

Alane Boyd (03:29)
Mm-hmm.

Micah (03:42)
I'm gonna make a judgment call about what I'm gonna pull out of what I saved before and insert it into this conversation. And it could be related, it could be completely unrelated. But because it's making a judgment call on both sides and AI isn't really that good at judgment, you can probably see the issue here.

Alane Boyd (04:02)
Yeah, and there's only so much context that the AI models can hold on to. So it's got to pick and choose what it's remembering about you.

Micah (04:10)
Yeah. I mean, that's a really good point because if we think about our memory, right, it's contained in our brain, I think. You know, that's the theory. And then at some point we pull stuff out and we're like, oh yeah, I remember when I got married or when I had my first kid. And contextually, we're making judgment calls about that, but that's what we're good at as people. With AI, if it stores a ton of information about you, it can't load

Alane Boyd (04:16)
Mm.

Mm-hmm.

Micah (04:36)
everything it ever stored about you and then use that in the chat, it blows out its context window.

Alane Boyd (04:44)
We have other episodes about context windows. It's when you have used it for a long time, it it cannot hold on to everything. And even in the moment that you're using it, we frame context windows easily for do you ever think that AI has lost the script on what you're talking it to and it's just spiraling and you're going, What the hell is going on?

And we're blaming it on the AI model, but it's really a user problem because we are expecting it to do something that it just cannot do right now.

Micah (05:05)
Yeah.

Alane Boyd (05:14)
All right, so let's go through the five issues or things that you're gonna run into if you're trying to use it from the memory that it's creating for you.

Micah (05:24)
Love it.

Alane Boyd (05:25)
So the first one, Micah, you already kind of talked about this, we don't have to stay on it, but it's that you're you're relying on it to make those judgment calls on what it's gonna save about you or the work that you're doing.

Micah (05:37)
Yeah, basically you have no control over what it saves and when it pulls that memory out to use it. Take that with a grain of salt because yes, you can go into the settings. Yes, you can delete, pieces of the memory that it's saved about you in the profiles that it's saved about you. But I'll tell you what, that is a pain in the ass. Nobody's gonna go in and do that.

Alane Boyd (06:03)
And you know, people will be like, well, I created a custom project or something like that where I uploaded files. Okay, great. That would be the next step above because you're telling it the context to go back to. And I want to argue a point on that your ChatGPT or your Claude in a custom project is not your single source of truth for your company. So you are uploading documents that are should be saved somewhere else in a shared knowledge base because you're

Micah (06:23)
Yes.

Alane Boyd (06:29)
most likely not the only person in your company or further gonna grow your company so you're a party of two. So

Micah (06:37)
Yeah,

I think that's such a great point, Alane, is the memories that it's saving are individual memories, not shared information. In a company aspect, that's really, important to have the shared aspect. The other thing that I want to point out is we're talking about two separate things, kind of. So if you're uploading files to a project, that's context for it to do its job that you're asking it to do.

Alane Boyd (06:54)
Mm-hmm.

Micah (07:01)
memories add to the context so it becomes the same thing. I guess it's all the same. Memories are context.

Alane Boyd (07:08)
I like to call it fake memories in this context. All right. So you touched on this group with the first one is that you don't have the control over what gets saved.

Micah (07:10)
Okay. Yeah. That works.

Alane Boyd (07:22)
So they're AI, it's a judgment call. Then it's not like you can go in and say, "hey, don't look at this again." And it sometimes will listen to you, but it may not look at that again for the next time it pulls the fake memory.

All right, which leads to bad memories are invisible.

Micah (07:40)
Ooh, that sounds scary.

Alane Boyd (07:42)
We don't always get things right the first time. A lot of times we're using things to brainstorm in there. Having somebody that can help piece things together, kind of do some gut checks on things. It's really kind of a first draft, that we're going to another team member to talk through further, but it's a a planning stage. Well, maybe not everything that you thought about ended up being part of what gets done. Well, it's gonna hold on to those.

Micah (08:06)
Yeah.

Alane Boyd (08:08)
brainstorms those things that you did and it may not be something that you wanted to access in the future.

Micah (08:15)
Yeah, it goes back to that judgment call. It might not be able to infer the context well enough to know that this definitely should not be saved as a memory. And then it creates a memory, fake memory. Sorry, Alane And then that is now kind of the truth moving forward. So the narrative of your profile, who you are and what you do and what you've done in the past can be completely wrong.

Alane Boyd (08:40)
Yes. And this leads me to the next one, which is it doesn't understand boundaries of context. So you might be using it personal and business-wise, it doesn't understand the difference in that. So it's going to leak that context between them. But I think more importantly than personal versus business is client versus client.

Micah (09:02)
Ooh, dang. That is huge, right? I'm just gonna say this out loud, so we can kind of talk through this. But the example would be I'm working on client A work. It saves a memory and then that thinks it's the truth, okay, with what we do as client A. But then I'm gonna go go move on to client B work. And suddenly it's pulling memories from client A that is

Alane Boyd (09:17)
Mm-hmm.

Micah (09:27)
changing the output of my client B work. Yeah, that sounds terrible.

Alane Boyd (09:31)
Yes. Really people kind of gloss over this, like, it's not that big of a deal. And I argue that it is because we're using AI to save us time, not just sit there and argue with AI. So even though yes, it's our assistant and it makes things easier, it's only making things easier if it gets it right. If not, we're circling with it.

Micah (09:50)
Yeah. And let me take that one step further,

real quick, Alane, because let's say you're rolling this out to a team and the team members don't understand, or you, as the person rolling this out, doesn't understand how this works, or the potential issues that this can cause. So you don't mention it to your team. So there's no awareness. And then all of a sudden your team members are going, Well, this sucks. This doesn't work. This is so dumb. It's easier to do it myself. And you went from

having a tool that could actually save your team members time to just being frustrated and not being able to get the adoption. And it's not the tech's fault. It's that the understanding of the underlying functionality around this. And this is why we're creating this episode, is because this is so important.

Alane Boyd (10:38)
It never fails because when we do an AI leadership workshop with a company, we always do a pre-workshop survey for the members that are gonna be in there. And always, there's always one that says, I tried using AI and it got it wrong and now I'm scared or it was a waste of my time. And I'm always like, my gosh, it wasn't the AI's fault. You didn't give it context, you didn't give it the right information. It still has to have something to work from.

Or you're just pulling from a general knowledge base, essentially, that AI has created on its own.

Micah (11:13)
Yeah. And and I would go so far to say, and this is a little off topic for the memory, but if you're relying on AI to make judgment calls in any form or fashion, then you're setting yourself up for failure working with AI. So the more that you can control, and the way that memory ties into that is if the automatic stuff is relying on judgment in and judgment out, already you can probably tell not a good idea.

But how do you then control it? So we're gonna touch on that in a second.

Alane Boyd (11:43)
Yes. So the last one of our five, and I've got a bonus one that I'm gonna add on, so maybe six, but is deleting a memory is harder than saving it. Because we use Obsidian and I love seeing how things are connected using Obsidian just with our internal knowledge base. It kind of helps paint a picture of how many things are intertwined with AI. And I don't know if those are

Micah (11:55)
⁓

Alane Boyd (12:09)
at all even connected, but it helps me understand, well, just because you try to delete it from one area of AI, it doesn't mean that it's not going to try pull in a memory from some other time and it's going to have it in there.

Micah (12:23)
Yeah. And who wants to sit there and read all the things that it saved about you just to surgically remove the one thing that could be causing problems and potentially missing other things that could be causing problems because of the way that it was worded or again, the judgment calls.

Alane Boyd (12:28)
Yeah.

Yeah. And so this is my bonus one. And if you are listening to this episode and you're just like Alane and Micah, who gives a shit? Like y'all are just like talking about something that really does not affect me. Okay, fine. This one should absolutely affect you because if you are depending on a memory for AI in ChatGPT or Claude or whatever you're using, that is personal to you.

Micah (12:50)
Yeah.

Alane Boyd (13:07)
It has nothing to do with anyone else in your company. And you are relying on every single person to use AI the exact same way as you and fed it the exact same information. And I guarantee you that is not going to happen. So every person is going to have a different result.

Micah (13:26)
Oh yeah, I don't even know where to start with this one, honestly, because if you have ever tried to scale a team or a business and leverage data, I guarantee you have said this phrase, source of truth. And we've had full episodes on this relating to AI. But essentially what you're saying, Alane, and and the the image that you're putting into my head is that if I were to roll this out to just say 10 people.

Alane Boyd (13:42)
Okay.

Mm-hmm.

Micah (13:54)
That

means everybody's memories are going to be different, even if they only use it professionally. I'm going to have 10 different sources of truth. And if they gave it the wrong context at some time or said something and they weren't quite sure and they were just asking questions, and then AI decided, hey, that should be a memory. Now their source of truth is off compared to everybody else's. And the outputs of everybody's work could potentially be.

Alane Boyd (14:00)
Mm-hmm.

Micah (14:24)
slightly different and shifted and changed because of this.

Alane Boyd (14:28)
This one gets under my skin the most because I know how important a single source of truth is when you're trying to scale a company because you want everybody working from the latest version of whatever it is that they need. And I notice that this becomes people start paying attention now with AI because they want they want their employees using it correctly. And so they start noticing like,

well, I've got employees saving things over here and doing things over here. And they notice how ineff not inefficient, but just how how people are operating differently. And now they're they're like, oh my gosh. And they think that it's AI's fault. I'm like, well, it was broken before that. It's just you're paying attention.

Micah (15:03)
Mm-hmm.

Yeah, it it illuminates it. It shows like, wait a second, this actually isn't what you thought was happening.

Alane Boyd (15:13)
Mm-hmm.

Yes. So we have a solution to this and we have another episode where we dive into this, but we wanted to frame it this way because okay. Okay.

Micah (15:24)
So Alane, hold on. Before

we tell her the solution, I want to say something that I think might, I don't know if it's going to be shocking. This is not controversial by any means. But we obviously use Claude, we use Claude Desktop. We use Cloud Cowork. We've made a ton of episodes about this. It's changed our business, how we operate, all that kind of stuff. I don't even know if you know this, Alane. At the organization level in Claude, we've shut off all memory.

Alane Boyd (15:34)
Yeah.

Micah (15:53)
There is no automatic memory being saved in our Claude account. Revolutionary, right? Yeah. Exactly. Nope nobody came back and said, well, it's not saving my memories.

Alane Boyd (15:59)
Whoa, I didn't know you did that. And you know what? I don't care. So thank you.

Yeah. Wow. So yeah, I didn't know you did that. And I will say that's because we have really focused on how we operate out of a folder system, going back to our single source of truth of where we already save documents and really focusing and even for ourselves, we uncovered some breaks in our process.

Micah (16:24)
Yes.

Alane Boyd (16:29)
You know what? We actually need to make sure we're saving things here and keeping files here because we notice, hey, we actually had some ways that we were doing it differently, either between teams or even between people working on a project.

Micah (16:43)
Yeah. So I want to say this only works with Claude Code, Claude Cowork, OpenAI Codex. It needs to be an AI system that can access the files on your computer because that's the feature that allows this magic to happen. Because if we think about what the automatic memories are, it's text snippets saved in the cloud. Fine. Well, what if text snippets are text snippets? So

If we say a memory is just a batch of text, we can say, let's control what the memories are by making those files. It could be a sentence, it could be a paragraph, it could be a whole brand guideline document. But by saving them in files and then using Codex or Cowork then Claude can load in those files, aka memories, when it needs based on the

task or the request that you're giving it. And all of a sudden you can completely curate the entire memory system, quote unquote memory, that your AI is using. And not only that, it solves the problem of the source of truth because you can have it synced on your Google Drive, synced on your Dropbox, whatever it is. And now you have one source of truth for the brain or the memory files that Claude or OpenAI is using forever.

Alane Boyd (17:45)
Mm-hmm.

Micah (18:04)
Everybody on your team.

Alane Boyd (18:06)
Yeah, and we're huge advocates of cloud storage that can sync to your computer that way that Cowork or Codex can talk to it. So this is the solution, just to be clear, that it's working from your company file system and accessing the files that you want it to see. And I mean, Micah, I love the brand one. it's just incredible because

not having to re-input things and it getting right the first time, it's not just a time saver, but just mental relief saver that I can just get it done the first time. Yeah. And you know, one of the things that I think about too, so like so many companies do RFPs or and and I shouldn't even say RFPs, like they do these large contracts.

Micah (18:37)
Sanity saver. We're gonna have to coin that phrase.

Alane Boyd (18:55)
That later have addendums to them. Well, you want to be, you want your team to be accessing the latest addendum to have the most correct information. Well, if you're working from a fake memory and you're uploading those files to projects individually, you're relying on that person to make sure that it has the latest version. Well, if you're working instead from your file system that you're already saving those addendums to, and your Claude.md file says,

In the instructions, hey, use the latest addendum when reviewing contract information, then that's already been said. The then Claude Cowork knows to access that file whenever you're using it.

Micah (19:39)
Love it. So the bottom line that this whole thing comes down to is if you want to have better outputs from your AI, memory is context. And all memories are is files with text in them, essentially. And you can organize it into folders, but you have to use tools like Codex or Cowork or Code to be able to access those files. If you're just using chat, old school now.

And you need to start looking at, especially from a business perspective, if you're using AI, look at Codex, look at Cowork, look at Code, look at these tools that actually provide the ability and then know how this stuff works so that you can make the right decision around the file structure, the folder structure, the memories, giving your team the heads up, answering these questions for your team for the rollout so you get the adoption. So you actually get the benefits of these tools, not just sitting here going, well,

didn't really work for me, so I'm just gonna make this spreadsheet myself.

Alane Boyd (20:43)
And we've got so many deep dives on Claude, setting up the file system. We've got a free webinar coming up. if you want to keep up with what we have going on that you can access, come join our Community. It's absolutely free. Micah does an AI News Brief every morning. This was actually one of the topics that he did a deep dive into for the AI News Brief. It's your.biggestgoal.ai, come and join the community for free.