Belief is built on credibility.
Not generated from a prompt.

Responding to "The Role of Organizational Memory in Scaling Enterprise AI," Forbes Tech Council, March 2026.

A recent article in Forbes, The Role Of Organizational Memory In Scaling Enterprise AI, asserts that to successfully scale enterprise AI, companies must treat their internal organizational memory (historical data, documentation, and operational context) as a strategic asset.

The article, in short: enterprises keep chasing AI success by buying better models and bigger datasets, but that's not where the problem is. AI struggles in a real company because it doesn't know how that company actually works — the decisions, the history, the lessons that live in documents nobody opens again. The author argues organizations should treat their institutional memory as a strategic asset: document decisions better, curate what matters instead of hoarding everything, review internal content regularly, and identify the key people who hold operational knowledge. Do that, he says, and the AI gets a cleaner signal, employees start to trust its output, and adoption spreads. His close: "AI doesn't replace organizational memory; it relies on it."

Here's my response.

The article gets the diagnosis right and the fix wrong.

The diagnosis is correct. AI doesn't fail in the enterprise because the model is weak. It fails because the model doesn't know how this particular company actually operates. Better models don't fix that. More data doesn't fix that. The missing piece is the organization's own memory — what it learned, what it tried, what happened. The author's line "AI doesn't replace organizational memory; it relies on it" is exactly right. I'd sign it.

Where I split from him is what he thinks organizational memory is, and what he thinks you should do about it.

He treats memory as a documentation problem. Write decisions down better. Curate the good stuff. Review your internal content on a schedule. Clean it up so the AI gets a clear signal. That sounds reasonable, and it's wrong in a way that matters.

The knowledge that's actually valuable is not the cleaned-up version. It's the raw version. It's the debrief someone sent at 11pm before anyone decided what the official story was. It's the transcript of two people who understood the problem actually working it out. It's the field note from the call, written before it got sanded down for the deck. That material is dense with the specifics of what really happened — and the specifics are the point, because the specifics are the one thing AI can't manufacture.

The moment you "curate" that, you strip out the very thing that made it worth keeping. Curation removes the particulars. It turns the debrief into a bullet point. You end up preserving the press-release version of your memory, which is the version that was already sanitized. So the standard advice — document more, curate, review — quietly destroys the asset it's trying to protect.

There's a second thing missing from his piece, and it's the more important one. He gets as far as "ground the AI in real operating context" and stops right before the actual answer: provenance. Context isn't enough. What makes a claim credible is that you can trace it back to a specific person who was there, who said it, and who can be held to it. That chain — this happened, this person saw it, here's the record — is what an AI can't fake. It can imitate the tone of something real. It cannot have been in the room. It cannot be accountable for what it says. That's the line that doesn't move, and it's the line everything should be built on.

So: right that memory is the constraint. Wrong that the fix is cleaner documentation. The fix is capturing the raw, specific, traceable evidence your organization is currently throwing away — and keeping the fingerprints on it, not polishing them off.

In real interaction, we've got built-in signals for what's credible. We've evolved to pick up and communicate credibility through cues in tone, hedging, and social proof references. AI has learned those signals cold.

What it can't produce is the residue of human presence. Historian Carlo Ginzburg called it dark proof: genuine knowledge surfaces through involuntary residue — things that couldn't have been staged because no one knew to stage them at the time. All those little idiosyncrasies are the proof of life that gets smoothed over by AI-generated content.

Most organizations have a strong story and almost no history. They have the line — the founding myth, the impact stat, the inflection year — but they cannot produce the artifact that proves the line. The board meeting that decided it. The customer who first articulated the need. The internal memo nobody filed.

I keep watching this play out in pitch rooms. The slide says "we serve over 40,000 students annually." The room nods. Then someone asks which student, in which classroom, last Tuesday. And the deck goes quiet, because the story was never built on a history.

This is the argument I'm still sharpening: belief moves on the verifiable. Not on the well-told. The well-told gets you in the room. The verifiable gets you the wire transfer.

The conversation, lightly edited:

CMO: So this is basically a content management system with extra steps.

Me: A CMS asks "where did we put the asset?" Idiograph asks "what actually happened, who witnessed it, and where is the residue?" The first is a filing problem. The second is an epistemology problem.

CMO: Give me the version I can repeat to my CEO.

Me: Your CMS stores what you said about yourself. Idiograph stores what is true about yourself, and lets every downstream system — writers, agents, decks, donor portals — draw from the same verified source. One is a folder. The other is a foundation.