Published August 19, 2026
For months, the default has been enough. I hand any model (Claude, GPT, DeepSeek) a memory that remembers facts about me and my work, and it works. That was the whole point of building EIDARA: to stop re-explaining myself every session.
But I’ve hit a ceiling. Remembering facts is now the minimum. To go further, I need the system to hold procedures too: reusable, well-built ways of doing things that I can hand to any model. The field has settled on calling them skills. When I went to add them, I ran into a question I’d been avoiding. A memory system looks like one thing, but it holds three different kinds of memory, and mixing them up is what creates the mess.
The three layers behave nothing alike.
The first is what I know, which means facts. A supplier raised prices. A course starts in October. Someone changed jobs. Facts arrive constantly, loose and out of order, in the middle of a conversation. There are too many of them, and they change too fast, to route each one past a human, so the AI curates them itself. It has to.
The second is who I am, which means my agents. An agent here is a way of working rather than a job title: “This is how my bookkeeping gets done: my spreadsheet, my bank, my judgment about what counts as a business expense.” These barely change, and when they do it’s because I decided to change how I work. The AI does not touch them. (An agent, to me, is just a document, but it’s my document.)
The third is how something gets done, which means skills. A skill is a procedure: “This is how you close a month,” or “This is how you build an expense report.” Skills change rarely, and only when someone who knows better improves them.
One sentence separates the last two: you download procedures; you never download identity. A closing-the-books skill works for anyone on earth. An agent knows which spreadsheet is mine and which judgment calls are mine. The skill is generic and travels; the agent is mine and stays with me.
I’m not drawing this line for the sake of a taxonomy. It settles the question that matters most once you let an AI tend its own memory: who has permission to write.
Facts are a stream. If I have to approve each one, the system dies of friction. Identity and procedure are the opposite. If a local model “improves” them on its own every night, one day you open your way of working and don’t recognize it, and you can’t say when it stopped being yours. From the outside they are the same files, with completely different rules about who may edit them.
So how do I stop the AI from deciding one day to polish a procedure I never asked it to touch?
The weak answer is to write a rule: don’t edit the skills. That holds until someone rewrites a component, or a new model reads the rule a little differently. Rules depend on everyone reading and respecting them, forever.
The strong answer is to put the skills out of reach. The part of the system that tidies memory only ever looks inside one place. Anything outside it is invisible to that part; it isn’t forbidden, it’s somewhere the tidier never looks. I have evidence that this works: the system’s own code has lived right next door for months and has never been touched, with no rule preventing it. Skills go there too, beside the code, reachable by any AI that consults the library and out of reach of the one that tidies memory.
This doesn’t mean the AI is dangerous. The local model that sorts my facts is good at its job: it reads a hundred loose facts, decides where each one goes, every hour, indefinitely. I let it. I just don’t want it also weighing in on a procedure written by someone who knows more than it does. It’s division of labor. Nobody finds it strange that the intern files the archive and doesn’t rewrite the quality manual.
A library of procedures turns into a dump within three weeks. Everyone uploads their own thing, half-duplicated, with no format and no note on what it’s for. I’ve watched this happen in shared folders my whole working life. So who enforces the format when the thing uploading is a different AI every time?
Nobody has to, because the format lives in the door. Every AI reaches my memory through the same small set of moves (see what’s in the library, open one, add one, retire one, delete one for good), and each move carries its own instructions, read at the moment of connecting. It already works this way for facts: when a model goes to save one, the door tells it the rules before it can. Skills enter the same way. If a skill is missing what it needs, it doesn’t get in.
Three things follow from that. It doesn’t matter which AI it is, today’s or next year’s, because they all cross the same door and learn the same rules. Nobody has to be trained: if I change the format in the door, every model knows it on its next connection. And order stops depending on discipline, which has never worked in any shared folder I’ve seen.
I made one decision on purpose: adding a skill is always triggered by a person, never on the AI’s initiative. A skill is a set of instructions the AI will then follow, and a system that downloads instructions and starts obeying them on its own is exactly what I don’t want running while I sleep.
The second decision was subtler. I first designed a review queue where the AI proposes and I approve. Then I admitted I was never going to read those files, so a review button would have been theater. That is worse than no control, because it looks like some. So I moved the check to what I can actually judge. I don’t judge whether the text is well-formed; I judge whether I just said “save that as a procedure.” A machine checks the text for format and size, and above all for personal data or credentials, which is the mechanical version of if it only serves one person, it isn’t a skill. I check the moment. You approve the moment, not the text.
Skills are the next layer of EIDARA, and like everything else here they come from hitting a limit rather than from a feature list. Facts taught the system to remember. Agents taught it who it’s working for. Skills teach it how the work gets done, without ever letting it rewrite the how on its own.
I’m still building this. The part I haven’t proven yet is the door under real use: whether every model that connects reads its instructions and respects the format, or whether some skills slip through half-finished. That is what I’ll be watching over the next weeks.
Javier
Facts, agents and skills, each with its own rules about who writes, are what I’m building after EIDARA v2.