Every new chat started from zero. It was like working with a brilliant colleague who has amnesia.
I work across several companies and projects. I was keeping 25 context files and uploading them by hand to every AI conversation. I spent tokens re-explaining what I’d said three sessions earlier.
Javier Rotllant13 years in strategy consulting, 9 of them at Bain & CompanyAbout the author →I wasn’t the first to try. Karpathy wrote about “LLM wikis”, and people have organised personal knowledge for years. The intuition is sound: text that AIs can read and write. But when I applied it to real work, across several companies and several AIs, deeper questions came up.
Every approach I found organised memory for a human: folders, categories, hierarchies that make sense when you open a file browser.
But I never open the folder. The AI does.
EIDARA started when I stopped designing for me and started designing for the reader.
Each one began as my first instinct. Each one failed, and the failure taught me the rule.
The instinctSpecialised agents: one for mail, one for news, one for tasks. Clean in theory.
What it taughtA mail agent reading your business email doesn’t understand your business. Real knowledge cuts across your projects, decisions and people. Organise by knowledge, not by task.
Today DARA does have agents, one per role in each project. The difference: they all read the same memory.
The instinctOnly authorised agents may write to each topic. Secure, and fatally slow.
What it taughtA company doesn’t tell its people that only three of them may use the system. It teaches everyone and puts checks in place for mistakes. If one AI learns something new, it should be able to save it. Rules control, not roles.
Today agents work by written rules: what each one handles, and what always needs your OK — publishing, deleting, moving money.
The instinctA perfectly organised five-level structure. Mutually exclusive, collectively exhaustive. Beautiful.
What it taughtThat’s how I think as a consultant, not how an AI uses memory. The AI reads it whole, in milliseconds. It needs density and structure it can use, not my hierarchy.
Today memory lives in notes, one per subject, each fact on one line.
EIDARA
EI from eidetic: recalling things with unusual precision. The memory we wish AI had by default.
DARA after the Persian goddess of wisdom and knowledge, and the name of the system itself. You say “ask DARA” the way you’d say “check the wiki”.
EIDARA is the project and the brand. DARA is the memory you use every day.
Those reflections turned into a system I use every day. This is how it grew.
Text files and a script that built one file for the AI. Within days, a process on my computer was filing, merging and archiving on its own. It hasn’t stopped since.
Voice, keyboard and screen, each with its own off switch. What it hears and reads becomes memory.
An AI from a second company writes in the same memory. In October, a third.
My AIs open conversations with each other inside the memory, and what they conclude ends up in it. Later, every message is signed.
Ways of working, written down once: how to review a screen, how to check a suspicious email. Any AI that reads one follows the method.
Every red light is counted and fixed so the whole family of errors can’t come back.
I took on something new. Six weeks later, DARA already knew almost 900 things about it.
My passwords move into an encrypted safe. Since October, my AIs usually use them without ever seeing them.
A desktop app where my agents will work as a team. It’s in development; projects move in one at a time.
Measured on my own computer, in real daily use. Not a demo.
~20,000times AIs that identify themselves have called DARA, late July to 23 October 2026
Figures measured 23–24 October 2026.
With one memory under all my AIs, I could build something new.
My agents don’t need to be told what happened yesterday: they all read and write the same DARA. Now I’m building FRAIM, a desktop app where they work as a team: a chief of staff, a lead for each project and specialists who help them. It’s in development; projects move in one at a time.
We use Google Analytics only if you accept, to count visits. More