Where the LLM wiki pattern actually fits

LLM wikis are everywhere in my feed the last few weeks.

The idea: instead of letting an LLM re-derive answers from your raw notes every time you ask, you have it incrementally maintain a structured wiki on top of those notes. A persistent, compounding artifact instead of a one-shot retrieval.

I built it into my own working repo a few weeks ago. Raw transcripts and clipped articles go into one folder. The LLM curates a knowledge layer on top, with an index and a change log so I can see what shifted and why. When I ask a question, the LLM decides whether the exploration itself is novel enough to file as a new page, so my queries become a source of new knowledge.

At my scale, it works. One operator, a high bar on what gets ingested, slow-shifting source material. Customer call transcripts, industry reports, competitive research. Quality in, quality out.

Before you build this for a team, two cracks worth sitting with.

Team scale breaks the input bar: Humans don’t edit wiki pages in this setup. They curate what goes into raw and read what comes out. The risk is upstream. More contributors means a looser bar on what gets ingested, and half-finished docs, off-topic notes, and low-signal transcripts pollute the inputs. With one curator, quality in, quality out. With ten the risk is garbage in, garbage out.
Dynamic knowledge doesn’t compound: A page about “Q2 priorities” or “current state of project X” is stale the moment anything moves. Re-deriving it on every change is exactly the cost the wiki was supposed to eliminate. The compounding artifact only compounds when the underlying truth is stable enough to compound on.

The bigger question this points at is how AI agents actually use organizational knowledge. Most of that knowledge isn’t wiki-shaped. It lives in CRMs and project tools, has multi-author quality control problems, carries sensitivity tiers, and shifts week to week. The wiki pattern handles one slice of that well. It’s not the whole answer.

Don’t pick a pattern because it’s trending in your feed. Before you build, ask what part of your team’s knowledge problem this actually solves.

If you run marketing or analytics: which of your knowledge sources are stable enough to compound, and which are alive enough that any synthesis is stale by Friday?

#AI #KnowledgeManagement #AIStrategy #LLM

First comment:

The pattern was popularized by Andrej Karpathy in this gist a few weeks back. Worth a read if you want to see the original sketch before deciding whether it fits your team: https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f