Published August 8, 2026
I open a brand-new chat with an AI I’ve never spoken to, type “hi,” and it already knows who I am, what I’m working on, and what I decided last week. No preamble, no re-explaining myself. The first time that happened it was uncanny. Now it’s just how I work.
The thing behind it is a persistent memory I’ve been building and living with for about four months. I’ve written before about how it’s put together — the compiling, the senses, the audits. This post isn’t that. This is the honest answer to the question people keep asking me: what do you actually use it for? Eight things, roughly in order of how much they’ve changed my day.
The memory lives on my own machine — nothing goes to the cloud. It keeps what matters about me and my work as small, self-contained facts and notes, and compiles them into a compact context file that any AI can read the moment it connects, through an open standard called MCP — the protocol that lets a model talk to outside tools. The AI never writes to the memory directly: a deterministic step decides what’s worth keeping, so the model suggests and the system decides. Right now it holds around 1,200 remembered facts across some seventy notes and about twenty reusable agent setups, compiled into roughly 8,000 tokens that any model reads the moment it connects — down from about 18,000 three weeks ago, after a pass that made the memory denser instead of bigger. That’s the whole idea. Here’s what it buys me.
Before, every new conversation started with ten or fifteen minutes of me re-explaining everything — who I am, my projects, the context — and I’d always forget something. Now, any model I connect — Claude, ChatGPT, DeepSeek — gets the compiled context on its very first call. Setup time: none. And they all share the same picture of me, so I’m not maintaining five slightly different versions of myself across five tools. The uncanny moment from the intro is this one. It stops feeling uncanny fast — and then you can’t go back.
Before, for a hard decision I’d ask three models the same question in three separate windows and merge their answers in my head, losing exactly the nuance where they disagreed. Now I open a shared thread where each model reads what the others said and then sharpens or pushes back. At the end I get a synthesis with the decisions and the disagreements spelled out. A few days ago I asked one agent to synthesize the operating playbook for a small internal app I’d built; it read the context, wrote the summary, I approved it, and it folded itself into the memory — no copy-paste. The deliberation happens in a scratch space that never touches the permanent memory; only the part I approve gets kept. I wrote about that table on its own.
Before, for my notes to contain something, I had to sit down and write it. Ninety percent of the time I didn’t bother; the other ten I forgot. Now it has three senses — it reads my clipboard, glances at what’s on my screen, and transcribes what I dictate — and a filter throws out the noise so the memory doesn’t fill with junk. I once read an email with a detail about a supplier and did nothing about it. The next day I asked what they’d quoted, and it had it. I never told it to remember. (It also has an off switch, which matters more than the senses — but that’s another story.)
Before, for anything repetitive — a review, a daily briefing — I re-explained the context every time. Now I keep around twenty reusable agent setups, each carrying its own context, rules, and references. I say the trigger and it starts with everything already loaded. I used to argue against “specialized agents,” and I still do — these aren’t that. They’re saved context, not narrowed roles: each one has the full picture, not a sliver of it. My morning-briefing setup took six rounds to get into the exact shape I wanted. That tuning is saved. I don’t lose it and re-derive it every time.
Before, in a folder-based notes app, I had to decide every single time where to put a thing and which silo to search. Now the memory is split into zones — personal, business, shared, system — with automatic gates. A business fact never leaks into a personal thread; a search crosses zones only where I’ve allowed it. I stopped thinking about filing entirely. The boundary is enforced by the system, not by my discipline — which means it actually holds.
Before, my old system just grew — duplicates, orphans, broken links — with nobody watching it. Now two auditors watch it: one runs daily from the inside, and one runs from outside the system with a deliberately critical eye. When something rots, they tell me. The last time the outside one graded the system it found debts the dashboard wasn’t showing; more recently it flagged a note that had quietly become 89% noise, and I cleaned it and added a guard so it can’t happen again. A memory that edits itself needs an outside eye, or it drifts into its own blind spots and calls it done.
Before, “what did I decide about this three months ago?” meant half an hour digging through old chats and email, and usually a half-answer. Now I search the memory and it returns the facts in time order, each one tagged with where it came from and when. It reasons about time — I can ask what was true as of a date, not just what’s true now. Seconds, with provenance. This is the difference between a pile of logs and something that actually remembers.
Before, twenty minutes each morning went to reading news, email, and my task list before I could actually start. Now a morning brief pulls together my calendar, my prioritized tasks, what the senses picked up the day before, and anything waiting on me — one page I open with a click. I start the day oriented instead of spending the first half-hour assembling the orientation.
Four months in, a few things have settled.
The difference between “notes” and “memory” is that memory decides what to forget. I don’t catalog; it curates. That’s the whole reason it stays useful instead of turning into a junk drawer I’m afraid to open.
The multi-AI threads turned out to be a new muscle. Deliberating before writing changes the quality of what gets kept — not just how the decision feels in the moment.
And the outside audit is the non-obvious ingredient. Any system that edits itself will, without an external eye, wander into a comfortable local minimum and declare victory. The critic that runs from outside is what keeps it honest.
None of this is finished. When something breaks — and it does — there’s a loop that catches it and closes it with a structural fix, usually within hours. But that’s a different post. This one is just the plain answer: these eight things, every day. Every week there’s a little more it quietly takes off my plate.
— Javier
EIDARA v2 is free. SUPER DARA is what comes next.