We're Training AI to Remember. We're Forgetting Our Own.

Context windows are now measured in millions of tokens. The leading models in 2026 can hold the equivalent of six or seven novels worth of your organisation’s decisions, reasoning, and history in a single session.

That sounds like progress. I think it’s a warning sign we’re misreading.

In this series, I’ve written about Cognitive Debt — decisions AI makes that nobody understands — and Retained Knowledge — the institutional memory that quietly disappears as humans step back from the work.

This piece connects both.

We built AI to extend our memory. We didn’t ask what we’d stop doing as a result.

There’s a concept called cognitive offloading — using external tools to reduce mental load. Writing a list. Using GPS. Completely normal.

The risk isn’t offloading. The risk is dependency — when the scaffold becomes structural and can’t be removed.

Research is now putting numbers on this. An Anthropic randomised trial found developers using AI scored 17% lower on comprehension tests than those who worked through problems themselves — with the biggest gap in debugging. An MIT EEG study found AI users had the lowest brain engagement of any group tested. They got the output. They lost the comprehension.

Multiply that across an enterprise. Sprint by sprint. Prompt by prompt.

Here’s the part that doesn’t get discussed enough.

The model accumulating your organisation’s context — decisions, exceptions, reasoning — becomes very hard to replace. Not because it’s the best model. Because it knows too much that lives nowhere else.

A cheaper model appears. A more capable one. Doesn’t matter. The switching cost isn’t technical. It’s contextual. The accumulated institutional understanding sits in weights you don’t own, on infrastructure you don’t control.

This is a new kind of lock-in. And most enterprises are walking straight into it.

So what’s the fix?

Not more documentation. Not RAG over existing docs. Not change logs written by the same AI that made the decisions. All of those capture outputs. None capture reasoning.

What we need is infrastructure that records not just what was decided, but why. What context applied. What precedent was set. Some are calling this a context graph — a living map of decisions and their interdependencies.

Right direction. But we don’t have it yet at enterprise scale.

Human memory has three layers: episodic (how things happened), semantic (what we know), and procedural (how we do things). Current AI knowledge tools capture only the semantic layer — facts and outputs.

The episodic and procedural layers — the how and the why — are still disappearing.

That’s what I want to explore next.

We’re training AI to hold more of what we used to carry ourselves.

The question isn’t whether the AI remembers.

It’s whether we still can.

---

Part three of a series on Cognitive Debt and Retained Knowledge. Next: the three-layer memory model and what enterprise knowledge infrastructure might actually need to look like.

#CognitiveDebt #RetainedKnowledge #EnterpriseAI #ContextGraph #InstitutionalMemory #AgenticAI #FutureOfWork