Three reflections that shaped how DARA thinks about memory.
The idea of using markdown files as persistent AI memory isn't new. Karpathy published "LLM Wikis," and the Obsidian/PARA community has been thinking about structured knowledge for years.
The core intuition is solid: text files that AIs can read and write. But when Javier Rotllant — 13 years in strategy consulting, 9 at Bain & Company — tried applying it to a real workflow across four businesses and multiple AI platforms, a series of deeper questions surfaced that needed different answers.
I manage multiple businesses — food & beverage, hospitality, technology consulting. I was maintaining 25 markdown context files, uploading them manually to every AI conversation. Burning tokens re-explaining what I'd already said three sessions ago.
Every new chat starts from zero. It's like having a brilliant colleague with amnesia. The question wasn't just "how do you give AI memory?" — it was "how do you design that memory so it actually works with how AI reads, not how humans browse?"
Every approach I found — PARA, wiki-style docs, hub files — shared the same assumption: organize memory for human navigation. Folders, categories, hierarchies that make sense when you open Finder.
But I never open the folder. The AI does. EIDARA was born when I stopped designing for the human and started designing for the reader.
DARA didn't arrive as a design. It crystallized after watching three natural instincts fail — and understanding why.
The first instinct: create specialized agents. A mail agent, a news agent, a task agent — each expert in their domain. Clean in theory.
But a mail agent reading your business emails doesn't understand your business. It processes format, not meaning. Real knowledge is cross-cutting — it lives at the intersection of your projects, your decisions, your context. Specialized agents can't see those connections.
Don't organize by task. Organize by knowledge. A domain expert reading their own emails understands everything. A task expert reading anyone's emails understands nothing.
The second instinct: only authorized agents can write to specific topics. Clean, secure, controlled. But fatally slow.
Think about it like a company. You don't tell employees "only 3 of you can use the system." You teach everyone how to use it and put controls in place for when they make mistakes. Otherwise it becomes a bottleneck — or, if you try to funnel all information through one point, you get overwhelmed with noise and nothing moves.
If new information surfaces in a DeepSeek conversation, why shouldn't DeepSeek save it? If Claude spots a correction, why wait for a "designated writer"?
Let everyone write. Let the compiler enforce quality. Rules control, not roles.
The third and deepest reflection. We studied PARA, wiki-style documents, tags vs folders. Built a 5-level structure: Projects, Areas, Resources, Agents, Archive. Perfectly organized. MECE. Beautiful.
And then I realized: as a consultant, this is how I always think and organize my thoughts. Categories, hierarchies, mutually exclusive structures. However —
It doesn't make sense to tell an AI: structure your memory in a way I can understand and browse, even if it complicates your life and is inefficient for how you actually process information.
This is what most people designing AI memory are doing right now. They're building for human navigation — folders you can click through, categories that make sense when you open Finder. But the human never opens the folder. The AI reads it cold, in full, in milliseconds. It doesn't need your hierarchy. It needs density and structure it can parse.
The radical decision: eliminate all hierarchical folders. A flat document lake + a compiler that creates structure. That became DARA.
EI — from eidetic: the ability to recall information with extraordinary precision. The kind of memory we wish AI had by default.
DARA — from the Persian goddess of wisdom and knowledge. Also 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 what you interact with daily.
EIDARA is the system that emerged from watching every intuitive approach fail. A persistent, compiled memory that lets you have any conversation with full context — regardless of whether you're working with Claude, Gemini, or DeepSeek. One memory, every model, every session.
A 10-step pipeline validates, deduplicates, and checksums your memory. Every compile is git-committed. Nothing accumulates silently.
Claude, GPT, DeepSeek, Gemini — any AI that can read markdown can participate. Tested across 4 models with a 92.6% average score.
Errors are expected. The next AI that spots one fixes it automatically. The system improves with every interaction.
Markdown files + Python standard library. No database, no Docker, no API keys, no cloud. Your data never leaves your machine.