---
type: "Article"
title: "Why I built Lorekeep: an LLM wiki is fine until it is your business on the line"
description: "Karpathy published the pattern and left the human review optional. Four months of running it as the only source of truth for a working consultancy is what turned that option into a gate, and the gate into a plugin."
resource: "https://www.laurelinlabs.com/blog/why-i-built-lorekeep"
tags: ["ai-knowledge-base-governance", "aeo-guardrails-and-agentic-search"]
generated: { by: "laurelin-site-build/2026-09-02", at: "2026-09-02T12:00:00+00:00" }
status: "stable"
stale_after: "2027-03-02T00:00:00+00:00"
---
# Why I built Lorekeep: an LLM wiki is fine until it is your business on the line

Published 2026-08-27. Author: [Piers Butler](/entities/piers-butler.md).

Lorekeep is the plugin I extracted from four months of running an LLM-maintained wiki as the only source of truth for a working consultancy. The pattern itself is Andrej Karpathy's: raw sources, an LLM-owned wiki, a schema file, and the operations that keep them current. What the pattern leaves optional is human review, and that is the part that breaks in a business. Wrong answers are survivable. Confident answers nobody can trace are not. So the human loop became a gate, the gate acquired a ledger, and the whole thing became something other people can install.

## FAQ

### Is Lorekeep just Karpathy's LLM wiki with extra steps?

It is the same pattern with the human review made compulsory rather than optional, plus the record-keeping that makes the review meaningful later. The gist is explicit that it describes an idea and not an implementation, and that on the business case humans in the loop are a possibility. Lorekeep is one opinionated instantiation of it, aimed at the case where the knowledge base has commercial consequences.

### What actually goes wrong if the model maintains canon unattended?

Three things, in rising order of cost. Contradictions get resolved silently, so the newest source wins whether or not it should. Claims that were true when captured stay in canon after they expire, because nothing distinguishes a fact from a fact with a shelf life. And nobody can answer why a rule is in canon, so it cannot be argued with or removed. The output still reads confidently, which is what makes it expensive.

### Why not just use a normal wiki, or a memory feature?

A normal wiki is written for people and dies of maintenance. Assistant memory features write inferences about you into context without asking, which is fine for preferences and wrong for commercial facts. Lorekeep sits between the two: a machine does the maintenance a human would abandon, and a human accepts every promotion an inference-based system would have made on its own.

### Is the methodology yours?

No, and the plugin says so. The underlying pattern is Karpathy's LLM wiki, itself in the tradition of Vannevar Bush's Memex. What is mine is the governance layer built on top after running it in anger: the promotion gate, the append-only oplog, ripple as proposal, index-first retrieval and the precedence order. Those are the open core, under Apache 2.0.

Hub: [Guardrails, Agentic & Future-proofing](/topics/guardrails-future.md)