Reflection

The MECE trap: how Bain taught me to think, but not to remember.

Published May 11, 2026

The consulting reflex

MECE (Mutually Exclusive, Collectively Exhaustive) is drilled into consultants from day one. You structure your analysis into buckets that don't overlap and that cover everything: issue trees, frameworks, waterfall charts. It's how you solve complex problems at places like Bain & Company, and after years there, it's how I solve everything.

So when I started designing an AI memory system, my first instinct was predictable: build a clean MECE folder structure. Projects, Areas, Resources, Agents, Archive. Exhaustive and non-overlapping. I was proud of it, and it made perfect sense.

To me.

How I remember, as opposed to how I analyze

A few days in, I asked myself the question that reframed the whole design: "How do I actually remember things?"

I meant remembering, not analyzing.

My answer was "by association." I remember what's relevant because I use it often. I remember connections between topics, not their position in a hierarchy. My memory works like a web of things that feel related, and nothing like a filing cabinet.

An LLM works the same way. It doesn't care about your folder structure. It reads context, finds relevance, and responds.

"That's how I think, but not how I remember."

MECE is a thinking tool, and a very good one for breaking down a problem. As a memory tool it fails. Forcing AI memory into MECE folders is like organizing your bookshelf by spine color: it looks tidy but doesn't help you find anything faster.

Deleting the folders

I eliminated all hierarchical folders. What's left is a flat lake of documents with two categories: NEURONS (what I know) and ENABLERS (how things work). A compiler reads everything and produces a single optimized file.

A Bain alumnus dropping MECE, probably the best-known framework in consulting, to design a memory system. I see the irony. The system works better for it.

Organize memory for retrieval, not for browsing

MECE describes how you think, and memory doesn't work that way, for humans or for AI. Most AI memory setups I see are still organized for a human browsing folders. If yours is, try one flat folder and a single compiled file for a week, and compare how often your AI finds the right context.

Javier

---

EIDARA is the open-source project that came from this realization. Code on GitHub, project website.