--- title: Why AI Forgets Your Previous Conversation — AI That Remembers (1) url: https://oosioo.com/en/p/%EA%B8%B0%EC%96%B5%ED%95%98%EB%8A%94-ai-1-%EC%96%B4%EC%A0%9C-%EC%B0%B0%EB%96%A1%EA%B0%99%EC%9D%B4-%EC%95%8C%EC%95%84%EB%93%A3%EB%8D%98-ai%EA%B0%80-%EC%98%A4%EB%8A%98-%EB%94%B4%EC%86%8C%EB%A6%AC%ED%95%98%EB%8A%94-%EC%9D%B4%EC%9C%A0 date: 2026-09-17T14:11:52+00:00 author: SYSOP summary: AI losing context isn't because you're bad at using it — it's because the way it stores things is built to collapse the moment you juggle multiple unrelated tasks. Here we trace the problem to its root. --- # Why AI Forgets Your Previous Conversation — AI That Remembers (1) It's the age of vibe coding, where you type what you feel and it just clicks into place. Other articles on this will tell you it's all just a click away. But anyone who has actually used AI coding tools for real work has run into this at least once. The AI that breezed through your project yesterday suddenly starts fumbling today, as if it's hearing about it for the first time, the moment you ask it to do something else. It forgets the approach you hashed out together after a long back-and-forth last week, breaks the rules you carefully wrote down, and keeps digging up and citing as evidence some random file you never even created. It's maddening. The short answer: it's not because your AI skills are lacking. The way most people use AI tools today is simply built to break down when you move between several unrelated tasks. And there are far more people juggling a bit of this and a bit of that than people who spend all day on a single task. ## The AI starts from zero every single time Most AI tools today can't carry context from a previous conversation or session into the next one once that conversation ends or the session changes. Open a new session, and it starts again from a blank slate. It's like a coworker who shows up every morning having completely forgotten what happened yesterday. They may seem capable enough, but if you have to explain everything from scratch every time, it gets frustrating fast. That's why modern tools have a device called a "rules file." Whatever a given tool calls its configuration file, the job is the same. You write down the rules to follow in the project folder, and the AI reads that file first thing whenever it starts working. It's like taping a list of office procedures to the desk of that forgetful coworker, something they'll read the moment they clock in. This approach works beautifully when you're digging deep into a single project. The rules are all in one place, and they're applied without fail every time you open that project. It's stable. ## The problem is the assumption of "one" Unfortunately, the work we do at a company rarely ends with a single project. In the morning you fix a bug in the shopping mall service, at lunch you tweak the plan for an internal dashboard, in the afternoon you analyze customer inquiry data, and in the evening you write a blog post. In a single day you move between four or five entirely different kinds of work. When the task is simple, you often don't even need AI for it. A rules file only works "inside this one project folder." The moment you step outside that folder, the context you built up in yesterday's project doesn't come with you. What's even more troublesome is something else entirely: what tone of voice you want, what principles you work by, what mistakes you told it last time not to repeat. This kind of information doesn't belong to any single project to begin with. So in the end it never gets written down anywhere, and you end up explaining it all over again every time. ## Writing more rules isn't the answer When people run into this problem, many respond by making the rules file thicker and thicker. Write this down, write that down, and before long it's hundreds of lines. But the thicker it gets, the worse two particular problems become. First, the AI has to read that entire thick rulebook every single time it starts a task. Most of it has nothing to do with the task at hand. The three lines you actually need right now get buried under a hundred lines of irrelevant rules. Even for people, doesn't a thicker manual make it easier to miss the one item that actually matters? Second, if you keep a separate rules file for every project, the shared content ends up duplicated across all of them. Things like your tone of voice or your principles get copy-pasted into file after file. Change one, and you have to hunt down and fix all the rest, and if you miss even one, the AI starts following different instructions depending on which project it's in. ![A scene showing that when all the rules are crammed into one file, the three lines needed right now get buried among unrelated rules](/uploads/c53669379bc39f6d.webp =720) ## The approach of scattering memory The method I've settled into after years of handling dozens of projects with a single AI is the exact opposite. I don't gather the rules into one file. Instead, I break them into small pieces, scatter them as "memories," and have the AI pull out only the fragments relevant to whatever it's working on at that moment. It's similar to human memory. We don't walk around with everything we know laid out in our heads at every moment. Only what's relevant to the situation in front of us naturally comes to mind. It would actually be strange if, while deep in concentration driving a simulated F1 track, you suddenly recalled a one-minute solo-cooking recipe from a show you watched last week. Since switching to this method, the problem of an AI that worked great yesterday spouting nonsense today has dropped noticeably. No matter how many projects I bounce between, the memories relevant to that task follow along on their own, and the information about me stays with it no matter what work I'm doing. In this series, I'll lay out the specific method over five parts. In the next installment, I'll take a closer look, from the angle of how storage itself works, at why cramming everything into one file is a bad idea.