Sattva

Your knowledge base is a city

Aerial view of a multi-level highway interchange winding between city blocks and autumn trees.
Photo by Denys Nevozhai on Unsplash

Every team’s knowledge base sprawls because it’s nobody’s job to steward it. The fix isn’t training people to do that job. It’s a knowledge system that stewards itself.

I was talking to an engineer friend the other day who helped me understand that software developers see their codebase like a city. Every time they push code, they’re using and changing shared infrastructure their colleagues rely on. So they need to be good citizens, good urban planners. And this is something engineers just know. It’s instilled in them from their first pull request.

Knowledge workers have a city too. It’s their knowledge base: all the institutional knowledge managed across various tools and data sources.

But unlike engineers, no one trains knowledge workers to be good stewards of it.

The reason for this is that when you’re a bad citizen in the codebase, the stakes are high. The platform goes down or velocity drops. To avoid those outcomes, engineers have developed many civic innovations over the decades. Version control, the system GitHub is built on, keeps a record of every change and who made it, so nothing is lost and anything can be undone. Pull requests work like building permits, putting every change through review. And automated tests act like inspections, flagging when something new threatens to break what’s already there.

These practices are highly compatible with today’s coding agents, like Claude Code and Codex. An agent can open its own pull request, run the tests before asking for review, and review a human colleague’s changes in turn. And because every change is on the record, anything an agent gets wrong is easy to undo.

In knowledge work, the stakes are just as real, but they’re invisible. Decisions are made with incomplete or stale information; repeatable tasks are started from scratch. But nothing goes down, and nobody gets blamed. Poor knowledge stewardship never crashes anything, it just makes everything crummier.

We know from years of trying that the fix is not making it someone’s job to manage the wiki, or enforcing standard operating procedures that everyone needs to follow. The fix is a system that stewards itself. And for the first time, that’s actually possible.

What does a self-stewarding knowledge system actually look like? That’s still being figured out, mostly by engineers building for themselves — Andrej Karpathy’s personal wiki is a great example. There are plenty of mechanisms now: RAG, semantic search, agentic memory. But stewardship goes beyond the mechanism. It’s the system being civic-minded about the knowledge base — fixing leaks, finding gaps, making connections, noticing redundancies, redrawing the map to keep it current. And what that looks like for a nonprofit team is different from what it looks like for an AI-native startup.

Nobody has settled this, but I’m convinced that the teams who solve this for themselves will be able to operate at a categorically higher level of effectiveness. That’s why Sattva exists!