The Re-Learning Tax

Ask an engineer the oldest question in any company — "which service owns billing, and what breaks if we touch it?" — and the answer usually lives in someone's head, three wikis, and a diagram that was true last spring.

Now every AI agent you run asks that question too. Silently, every session, at your expense.

Stateless by design

A language model holds whatever fits in its context window and nothing else. Between sessions it remembers nothing. This is not a defect — it is the design — but it means every session starts ignorant of your company and must rebuild its picture from whatever happens to reach its context.

You pay it every session

The rebuild is not free. You pay in tokens — context repeatedly re-assembled and re-sent. You pay in time — agents re-deriving what the organization already knows. And you pay in mistakes — because an agent that cannot find a fact does not stop; it guesses, plausibly and confidently, from training-data averages that are not your company.

Worst of all, each agent rebuilds a slightly different picture. Ask two agents the same question and you get two answers. An organization running on as many versions of the truth as it has context windows is not coordinated; it only looks that way on a good day.

Documentation rots; systems do not lie

The fix is not more documentation. Wikis go stale the day after they are written. But repositories, clusters, and manifests cannot drift from reality, because they are the reality. So that is the place to learn from: scanners walk the actual systems and propose facts — services, owners, dependencies, relationships — each grounded in a source you can open.

Curation is code review for facts

Raw extraction produces duplicates, conflicts, and noise — the same reason raw code needs review. A curator pipeline merges and deduplicates automatically; contested claims wait for a one-click human ruling. Every surviving fact carries its sources, verification state, and freshness. Not accumulated — curated.

Learned once, known everywhere

The moment a fact is verified it belongs to the whole workforce — cloud workers and enrolled developer assistants alike — served over MCP and APIs, injected into context at the moment of use. The question is answered once, by the systems themselves, and never becomes folklore again.

A static wiki depreciates from the day it is written. An org brain fed by scanners, curators, and working agents appreciates instead.

That is the end state of paying off the re-learning tax: every verified fact makes every future task slightly cheaper, faster, and safer — and your organization gets smarter every day it runs.