An LLM wiki is not a chatbot with memory

You ask your AI tool what the four sources on DORA you saved actually said. It answers. The answer is good.

Three weeks later a colleague asks the same thing. She gets a different answer — also good, but phrased differently, weighted differently, perhaps drawing on three of the four sources instead of all four. Nobody notices. There is nothing by which anyone could.

This is where two things that look alike from outside come apart.

An answer is an event. A page is an object.

A chatbot with memory produces answers. An answer is an event: it happens in a conversation, for one person, at one moment. Afterwards it survives at most in a transcript nobody opens.

An LLM wiki produces pages. A page is an object: it has an address, a current state, a history. Two people can mean the same page. You can link to it, put it in a deck, contradict it.

That sounds like a detail of presentation. It is the whole difference, and it has three consequences.

Consequence 1: a page can be visibly wrong

That is not a drawback. That is the point.

A wrong chat answer has scrolled away before anyone objects. And because the model produces the answer fresh each time, the same mistake reappears on the next question. You cannot fix it, because it is not written anywhere.

A wrong wiki page sits in one place. You can read it, dispute it, change it. You can look at the version history and see what it said before and who changed it. A mistake you can pin down is a mistake you can be rid of.

That does assume the page is checkable — that every claim names its source, and that the base does not quietly drift away from itself. We wrote about that elsewhere, because it is the less comfortable half of the same coin.

What "visibly wrong" means in practice

Take a page that says: "The reporting deadline is 72 hours", with a citation behind it. Three things are now possible that none of them are with a chat answer.

Someone opens the citation and finds it says 24 hours — the mistake is located, not merely suspected.

Someone changes the figure, and the history records that it previously read otherwise. Whoever has to defend the July memo can look up what they were relying on at the time.

And when the rule itself changes, the source changes — not your memory of it. A knowledge base can notice that. A transcript cannot.

"But my chatbot does cite sources"

That is the best objection, and it deserves an honest answer: citations in chat answers are a real improvement on free generation. They still do not close the gap.

A citation in an answer backs that one answer. It does not tell you whether the same question yesterday rested on the same sources. It does not tell you which sources the model saw and set aside. And it tells your colleague nothing, because her answer carries its own citations.

The difference is the durability of the evidence. On a page, the source list is part of the object: it changes when the page changes, and both are readable afterwards. In an answer, the citation is a property of the moment.

You feel this the second somebody asks: "Where did you get that?" With a page you point. With an answer you go looking through a transcript.

Consequence 2: a page holds still while you use it

Professional work is rarely finished in one sitting. You read about something in May, write a memo in July, and the follow-up question comes in October.

With a chat interface that means asking three times, getting three slightly different answers, and judging three times which one holds. The coherence of your work rests on you.

With a page it means the same page three times. If it changed, you see when and why. The coherence rests in the record.

For advisory work that is not a convenience but a precondition. A client comparing two statements of yours is not comparing two transcripts.

And it is where the opening example returns: that you and your colleague got different answers is harmless in itself. It becomes dangerous when both end up in client material — and nobody ever learns that they diverged. A shared record makes the divergence impossible. A chat makes it invisible.

Consequence 3: a page belongs to several people

A chat answer has exactly one addressee: the person who asked. A team relying on AI answers has as many versions of what it knows as it has people.

A page belongs to everyone who can see it. That sounds obvious, but it changes what is possible at all: someone can add what another person read. An engagement can have a page, instead of four people with four recollections.

How the two fit together

This is not an argument against chat. It is an argument against chat without a record.

The division of labour that works:

The wiki is the record. Distilled from what you actually read. Every claim with its source. Durable, correctable, shared.

The chat is how you query it. You still ask in plain language — the model just answers from your record rather than its own memory. In our case that runs over MCP: your AI tool reaches into the cited wiki, and the answer can say where it came from.

The difference shows in a single question you put to the result: can I send this to someone?

A chat answer you can copy. A page you can link — and the person at the other end sees the same page you do, with the same sources, in the same state.

Why the distinction matters now

"Memory" has become a selling point. Every major AI tool now retains something about you, and that is genuinely useful — it saves repetition.

But memory is not knowledge. Memory is what the tool retained about you. Knowledge is what you checked and recorded, and what is still there when you change tools.

That is the practical test I would apply: what remains when you switch vendors? With a chatbot that remembers: a transcript nobody reads. With a wiki: the pages, with their sources, exportable.

That is not the same product with a different interface. It is a different object.