Every enterprise AI pitch deck has a slide about “capturing institutional knowledge.” The story writes itself: senior people leave, they take decades of context with them, and if we’d just built the right system we could have preserved it. Knowledge management 2.0, now with embeddings.
The problem is that the most valuable knowledge in any organisation exists because it’s not written down.
Think about what actually makes a veteran operator effective. Sure, they know the processes — but so does anyone who reads the policy manual. What they really know is which compliance officer actually reads the submissions and which one rubber-stamps, which vendor relationship is fragile, why the last platform migration was quietly shelved, what the CEO said in the corridor after the board meeting. This is the knowledge that determines outcomes. And it lives in conversations that people would deny having.
Here’s the paradox: build a system that captures this, and people stop saying it. The value of dark knowledge is inseparable from its informality. The moment someone knows their offhand remark about a flaky data pipeline will be indexed, vectorised, and surfaced to an AI assistant next quarter, they stop making offhand remarks. You’ve instrumented the very behaviour you wanted to preserve.
This isn’t a technology problem. It’s a trust architecture problem wearing a technology costume.
Enterprise knowledge actually exists in three layers. The explicit layer — policies, procedures, documentation — is already handled. RAG over your SOPs works fine. The operational layer — preferences, patterns, “we tried that in 2019 and it broke” — is genuinely capturable, and this is where agent memory systems earn their keep. But the political layer, the informal layer, the layer where real decisions get made before the meeting where they get announced — that layer is adversarial to capture by design.
The consulting instinct is to sell a system for all three. The honest answer is that if the third layer is large — if a significant fraction of critical knowledge lives in conversations people actively avoid documenting — that’s a governance design problem, not a knowledge management problem. People hoard context because formal systems are too rigid, too surveilled, or too disconnected from how decisions actually get made.
The right question for a client isn’t “how do we capture tribal knowledge?” It’s “why does so much critical knowledge stay informal in the first place?” Fix that, and you need less memory infrastructure. Ignore it, and no amount of vector search will help.