Story How it works Results Download Contact

The story

How a Bain consultant's question broke every assumption about AI memory.

The reference — Karpathy got it almost right

Andrej Karpathy — former AI director at Tesla, OpenAI co-founder — published a paper on "LLM Wikis": markdown files that AI agents maintain as persistent memory.

The core idea is solid. But Karpathy, like everyone else, made one mistake.

The problem

Javier Rotllant manages multiple businesses — food & beverage, hospitality, technology consulting — with 25 markdown context files he uploads manually to every AI conversation.

Every new chat starts from zero. It's like having a brilliant employee with amnesia. The question: how do you give AIs persistent memory without heavy infrastructure, without vendor lock-in?

Karpathy's mistake (and everyone else's)

Karpathy, Tiago Forte (PARA), and the entire Obsidian community all make the same mistake: they design AI memory for human navigation.

EIDARA was born when Javier questioned that fundamental premise.

Three iterations

v1 — The org chart trap

First attempt: 6 specialized agents (mail agent, news agent, task agent) + a Librarian + a CEO supervisor. Javier killed it immediately.

His argument from Bain & Company: "A mail agent reading TFFC emails doesn't understand TFFC. A TFFC expert reading their own emails does. Don't organize by task. Organize by knowledge."

Task experts are useless. Domain experts are everything.

v2 — The roles trap

Second attempt: domain agents with write permissions. Only the TFFC agent writes TFFC memories. Clean, secure, organized. Javier killed it again.

"Having dedicated writer agents limits me. I want EVERYONE to write — any Cowork session, any Claude chat, any TypingMind. If new TFFC info comes up in a DeepSeek conversation, why can't DeepSeek save it?"

Rules control, not roles. The protocol is the security.

v3 — Think like an AI, not like a human

We studied PARA, Karpathy, hub documents, tags vs folders. A 5-level MECE structure: Projects, Areas, Resources, Agents, Archive. And then Javier said something that broke everything:

"Everyone else is wrong too. This is how humans think. But it's not right for AI. You already know how to manage your info. I ask you for it and you give it to me. Your info isn't in MECE structures, it's in complex models that measure frequencies and search. It doesn't make sense for me to tell you: have a memory like mine."

This is profound. Everyone — Karpathy included — designs AI memory thinking about how a HUMAN would navigate those files. But if the human never opens the folder, who are you organizing for?

The radical decision: eliminate ALL hierarchical folders. A flat document lake + a compiler that creates structure. That became DARA.

Why "EIDARA"

EIDARA is the project and the brand. DARA is the system. You say "ask DARA" the way you'd say "check the wiki."

The name was verified clean across Google, GitHub, PyPI, npm, and international trademark databases.

About the author

Javier Rotllant Miras

Former Bain & Company. Runs multiple businesses across F&B, hospitality, and technology consulting.

Built EIDARA because he needed it. Not an engineer — a PM with a problem and the stubbornness to solve it. The kind of person who tells an AI architect "everyone else is wrong, think differently" — and means it.

Build something useful, share it honestly, and see where it goes. The right people notice.

GitHub [email protected]