--- title: AI That Remembers (3) — One File, One Fact url: https://oosioo.com/en/p/%EA%B8%B0%EC%96%B5%ED%95%98%EB%8A%94-ai-3-%ED%95%98%EB%82%98%EC%9D%98-%ED%8C%8C%EC%9D%BC%EC%97%90-%ED%95%98%EB%82%98%EC%9D%98-%EC%82%AC%EC%8B%A4%EB%A7%8C-%EB%8B%B4%EB%8A%94%EB%8B%A4 date: 2026-09-20T11:43:15+00:00 author: SYSOP summary: Storing memories well matters less than storing them so they can be found. How to split them by kind, the one-file-one-fact rule, and the one-line description that decides whether they get recalled. --- # AI That Remembers (3) — One File, One Fact I said memory gets broken into small pieces and scattered around. This is where most people get it wrong. They think the goal is just to pile up as much as possible. What actually matters isn't piling up a lot — it's piling it up so that it can be found exactly when it's needed. Saving is easy. Recalling is hard. ## First, split by kind If every memory goes into one basket, things turn into a mess later. I split mine into four broad categories. - Who I am: my role, area of expertise, preferences. Information that should sit in the background no matter what task is at hand. - How to work: instructions and feedback I've been given. Do it this way, don't do that. Especially mistakes that were pointed out once already. - Reference material: addresses I look up often, document locations, config values — external information. - Work in progress: the goals and status of projects currently running. Context you can't get just by looking at code or logs. Splitting things this way makes management easier later. You can pull together just the feedback on working style, or sort out only the memories from a finished project. ## One fact per file The second principle is this: one file holds exactly one fact. If you just cram things in loosely and try it out, it works better than you'd expect. That's when the temptation hits to lump all the related content into one file. Doing that breaks two things. First, recall becomes imprecise. If ten topics are mixed into one file, the AI pulls out the whole file because of just one of them, and the other nine unrelated pieces come along for the ride. Second, updating gets harder. When one fact changes, fixing just that part in a file where several facts are tangled together turns out to be more of a hassle than it looks. With one file per fact, only that exact thing gets pulled out when needed, only that file needs opening when it changes, and only that file needs deleting when it turns out to be wrong. ## What decides recall is the 'one-line description' This is the most important part of this piece. At the top of every memory fragment, attach a one-line description of what this memory is about. That single line is both what goes into the index and what the AI uses to judge whether to pull this out right now. So how you write that one line decides whether recall succeeds or fails. Take an example. A bad description looks like this. > Note about payments A good description looks like this. > Rule of three retries on payment failure at the shopping mall, and the refund approval process A bad description only reacts to the word "payment." A good description spans concrete situations like retry, refund, approval. You imagine the context in which the AI would need this memory, and plant the words of that situation in advance. Think of it less as summarizing the content and more as writing down when it should be pulled out. ![Comparison of recall between a bad description that just says "payment note" and a good description packed with various situational words](/uploads/c5ea2efeb1b3f2d1.webp =720) ## Two small habits Lastly, two habits that have stuck after long use. One, always write dates as absolute dates. Expressions like "last week" or "yesterday" lose their reference point by the time the memory gets recalled later, turning them false. Writing "September 15, 2026" instead of "the third week of September" keeps the meaning intact even months later. And a lot of LLMs turn out to be weaker than expected at date lookups. I've always written the date together with the day of the week. Two, link memories to each other. When one memory relates to another, put in links pointing to each other. So that pulling out one naturally leads into related ones. Once scattered fragments are connected this way, they stop being scattered notes and become a single web of knowledge. In the next installment, I'll show how memory built up this way actually operates in a real day's work, following a day that bounces back and forth between tasks of different character.