Adam Sand explains why AI models like ChatGPT forget earlier conversation details: they have a fixed token limit called a context window. When it fills up, the model discards earlier information or shortens responses. He shows how prompt-engineering files can give AI persistent identity across separate chat sessions, and recommends leaving "breadcrumbs" to reconnect new conversations to old context.
Why does ChatGPT suddenly get lazy and forget everything you just told it? You didn't break the AI, you hit the context window. So you get a context window. So you have 200,000 tokens here. Now, right now you are at that much tokens. Now what happens is once that pops up and starts saying you're here, AI has a thing called context anxiety, which means that as it gets to this yellow mode, it starts going, ooh, I'm running out of context. I need to start getting shorter responses. I need to try and get to the end of it, like get to the point. It's like its attention span. And then once it gets into the red, it starts to do what's called auto-compaction. So it starts compacting the conversation. So what's your favorite movie?
Um, Inception.
Inception. You picked the hardest fucking movie. I'm sorry.
Think of another one.
Interstellar. Interstellar. Okay, Interstellar. So in the middle of that movie, it's gonna start forgetting the fact that when he left, his daughter was noticing her toys falling off the shelf, which is a huge part of the movie, but it only matters at the end because it's when he's hitting the books, right? So it only matters at the end of the movie. So if you hit the thing on the shelf and the toy falls off, you don't have that aha moment because when it was compacting the context, it was like, oh, that was just a weird thing that happened. A toy fell off a shelf. That's why it literally has to remind you in in the movie. So this is what I mean by context compaction. And so you have this context window and each session lives on as its own individual thing. It's like it has a 200 IQ, but it wakes up with amnesia.
This is a context window, right? So if I open another one, that would be a separate context window, correct?
Yes. But then it is only aware of what's in here. It is not aware of what's happening in here, because this is like writing a CD on Claude's computer. Like, this CD is only aware of itself. The next CD that you write is only able to access the same unified file set, or they might call it a runtime, right? So it's aware of this. So that's why we create a file called ClaudeMD, because ClaudeMD preloads at the top of this, every new session you ever open inside this folder, it looks at ClaudeMD to say, what the hell am I? And that's pre-written into Claude's brain. So it says, whenever you open up a new session, check if the place that you are contains a Claude.md. And now if you, you were using ChatGPT or many of the other tools, the default is agents.md. And if you were using this in Antigravity, right, because you could open this folder up in Antigravity and it would look at for Gemini.md. So we usually tell it to clone these files so that they're the same. So, if you fill up your context window, you have to make sure that by the time you get to here, you are creating something that is going to lead you back to here. You have to leave yourself a breadcrumb.
