Your knowledge base should be plain markdown

When people set out to build a knowledge base for their AI, the first instinct is to make it look serious. Polished PDFs, a wiki with nested pages, a slide deck for the strategy. It feels professional. It is also the wrong foundation, and it quietly works against you.

The knowledge your AI reads should be plain markdown: readable text with a little structure, stored as files. Boring on purpose. Here is why boring wins.

Machines read plain text, not layout

A PDF is a picture of a document. A slide is a canvas. To a human they look organized; to a machine they are a mess of positioned fragments where the reading order is a guess. Spreadsheets are worse: meaning lives in the layout, and layout does not survive the trip into an AI's context.

Markdown is the opposite. It is text with just enough structure, headings, lists, tables, that both a person and a model read it perfectly, the same way, every time.

If your AI has to reverse-engineer your formatting before it can read your knowledge, you have made its job harder, not easier.

This is not a small tax. The gap between "the model understood the document" and "the model half-guessed the document" is exactly the gap between a grounded answer and a confidently wrong one.

Plain text is knowledge you can actually own

The format question is really an ownership question. Plain files are yours in a way a proprietary format never is:

• You can read and correct them without a special app. When something is wrong, you fix a line.
• You can track every change with ordinary version history, so you always know what changed, when, and why.
• You are not locked in. The same markdown works with any AI, today's and next year's. Nothing is trapped inside one vendor's tool.

Pour your company into a pretty black box and you have a filing cabinet you cannot inspect. Keep it as plain files and you have knowledge you control. That distinction is the whole reason we insist on it.

"But it looks unprofessional"

It looks unprofessional the way a clean database looks unprofessional next to a colorful chart. The polish belongs on the output, the proposal, the deck, the email your AI generates from the knowledge. It does not belong on the source.

Keep the source plain and readable, and let the AI produce the polished artifact on demand. You get both: knowledge that machines read cleanly, and output that looks as good as you want. Trying to make the source itself beautiful just corrupts the part that needs to stay legible.

What this looks like in practice

Say your pricing lives in a file called pricing.md. It is plain text: a heading, a short paragraph on positioning, a list of tiers. Nothing fancy. When someone asks the AI to "draft a proposal for a mid-size client", the model reads that file cleanly, every tier, in order, and produces a polished proposal with the numbers right.

Now picture the same pricing trapped in a designed PDF. The model has to guess at the layout, and a guessed number in a proposal is how you lose a deal. The plain file is not the ugly version. It is the reliable one, and the polish moves to the output, where it belongs.

Common questions

What about diagrams and images?

Keep them, just do not make them the source of truth. A diagram can illustrate a document; it should not be the document. Put the meaning in text and use the image as support, so the AI is never forced to read a picture to understand your business.

Isn't a proper wiki better?

A wiki is fine for people to browse, but the content underneath should still be plain text you can export and version. The danger is a tool whose format you cannot get your knowledge back out of. If it is not portable, it is not really yours, which is the same ownership point behind deciding what to put in your company memory.

What do I actually write it in?

Any plain text editor works. A free markdown editor makes it nicer to look at, but the file is just text, which is the entire point. The tooling is covered in the practical Claude Code guide.

What about a spreadsheet of data?

Data that is genuinely tabular can stay tabular, exported to a simple text format the AI can read. What does not survive is meaning that lives in cell colors, merged headers, and layout tricks. If a human needs the formatting to understand it, the model will struggle too.

The takeaway

Resist the urge to make your knowledge base look impressive. Make it plain, readable, and yours. Markdown files are the least glamorous and most durable foundation you can give your AI, and everything else, search, memory, governance, gets easier when the source is clean text.

That principle is baked into memrelay: your company's memory is plain markdown you control, changed through review, and readable by any AI client. If you want to see how it fits into a working setup, start with the practical Claude Code guide.