Notes on knowledge work

Who actually maintains your LLM wiki?

Building an LLM wiki is the easy part. Keeping it true is hard: deleted sources still cited, sources that quietly change, two pages about the same thing. How we separated sorting, building and syncing — and what an agent must never do on its own.

Three days, one story: the audio roundup built from your morning reports

Every few days your morning reports become an audio briefing: one thread, at most three lines, each naming its source in the sentence, upcoming deadlines at the end. It needs no saved articles — and it sounds calm because we worked on the pauses and the music.

The morning report: what happened in your field, with evidence

Every morning at six, the most relevant articles of the last 24 hours are summarised from the articles themselves and written up as a short report: situation, key points, a section per theme, every sentence with its evidence. What goes into the wiki is your decision, article by article.

Who appears with whom? The entity graph on the Topic Map

Next to the concepts of your wiki, the Topic Map now shows who and what your bookmarks name: people, companies, products, regulations — linked when they appear in the same articles. Why cleaning matters more than drawing.

What is an LLM wiki? The basics

An LLM wiki is a wiki a language model writes and keeps from your sources: pages per concept, every paragraph with its evidence, every change as a version, disagreements between sources made visible. How it differs from search over a folder of files and from a chatbot with memory.

An LLM wiki is not a chatbot with memory

Both answer your question from your own material. The difference is what remains afterwards: an answer nobody can check again — or a page that stays, can be cited, and is allowed to be visibly wrong.

When your knowledge base is wrong about itself

A machine-maintained knowledge base has a failure mode a handwritten one does not: it can be wrong about itself and look entirely fine. What we found auditing our own wiki — and what follows from it.

Where did I read that? Every mention, in one place

DORA turns up in four things you read across three months. You remember two. Pith recognises the named things in everything you save — regulations, institutions, companies, people — and gathers every mention in one place.

Your reading, as one knowledge graph

You don't read in silos, so why store reading in them? Pith distils everything you save into one knowledge graph — and the adjacencies it surfaces are where the insight lives.

Every claim cites its source — and why that beats a chatbot that invents one

A chatbot will hand you a confident sentence with no idea where it came from. Pith hands you the same sentence with the saved source attached. For defensible work, that difference is the whole job.

Reading memory for your AI: Pith's hosted MCP server

Your AI assistant has no memory of what you've read. Pith's hosted MCP server gives it one — a cited reading memory it can query, with every claim linked back to its source.

The Founding Practice deal — fifty firms, locked for life.

Archive: our May 2026 launch offer of €25/seat for the first 50 firms. Pith has since retired seat subscriptions entirely — saving is free, and you pay per deliverable.