Published May 11, 2026
Every approach I found (PARA, Karpathy's LLM Wiki, Obsidian vaults, hub documents) shared the same assumption: organize your AI's memory the way you'd organize your own files, with folders, categories and hierarchies that make sense when you open Finder or Explorer.
After years at Bain & Company, I'm trained to think this way. MECE frameworks, issue trees and hierarchical structures come to me by instinct. My first prototype had a tidy folder tree with Projects, Areas, Resources, Agents and Archive. It was MECE-compliant and perfectly logical, and it was the wrong design.
When I asked myself how I actually remember things, the answer wasn't "in folders." It was "by association." I remember what's relevant because I use it often, and I remember the connections between topics, not their position in a hierarchy.
That realization became the center of the design: "That's how I think, but not how I remember."
My brain uses MECE to analyze problems, but it doesn't store memories that way. It stores them by relevance, frequency and connection. An LLM works much the same way: you give it context, it finds what's relevant, and it responds. It doesn't browse directories. It reads everything at once, in milliseconds, and finds what matters.
So who are you actually organizing for? The human who never opens the folder, or the machine that reads it all in one pass?
I eliminated all hierarchical folders and kept a flat lake of documents. A compiler reads every file and produces a single optimized output, which I call BRAIN.md.
There are two categories and no nesting. NEURONS hold what I know: projects, contexts and knowledge, one topic per file. ENABLERS hold how things work: agents, credentials and tools. That is the entire structure, with no subfolders and no categories within categories. The organization is created by the compiler at build time, not by me at write time.
Tens of thousands of characters of source material compress down to ~3.5K tokens of optimized context. Any AI reads it cold, in full, and has everything it needs. There's no folder navigation, no "where did I put that file?" and no hierarchy to maintain.
I've been running this daily across 4 businesses for weeks. It works because I stopped designing for myself and started designing for the reader, and the reader is the machine. If you keep notes for an AI, count how many of your folders the model ever needs to see as folders; in my case the answer was none.
Javier
EIDARA v2 compiles a flat folder of notes into one BRAIN.md any AI can read.
This is part of a series on building practical AI memory systems. EIDARA is the open-source project that came out of this approach. Code on GitHub, more on the website.