The AI wiki: stop re-reading, start compounding

Watch an assistant answer a question about your business and you will see something wasteful: it reads the same scattered documents it read yesterday, reconstructs the same understanding, answers, and then forgets all of it when the session ends. The reading was rework. The synthesis evaporated. An AI wiki fixes this with an old idea and a new librarian: write the understanding down once, keep it linked and current, and make everything, human or machine, read the summary before the source.

The shape of the idea has been circling for a while; Andrej Karpathy has sketched it as the wiki whose maintainer is a language model. The team writes and works as normal. The model does the librarianship: ingesting new material, updating the summary pages, linking concepts to each other, and flagging contradictions for a human to settle. Synthesis happens once, on ingestion, instead of on every single question.

Context windows are large now, but attention is not. Stuff a model's context with two hundred thousand tokens of raw documents and the answer quality degrades; the signal drowns in its own supporting material. A curated page that already contains the synthesis costs a fraction of the tokens and answers more accurately, because a human-reviewed distillation is denser in truth than any pile of originals.

Synthesis happens once, on ingestion, instead of on every single question.

• A sources folder. Anything new lands here: meeting notes, decisions, vendor docs, postmortems. Nobody formats anything.
• A librarian pass, on a schedule. An agent reads what is new, updates the relevant summary pages, adds cross-links, and opens a change for review rather than editing silently.
• Summary pages, one per concept. Short, current, dated, and linked to related pages, so a single retrieval brings connected understanding with it.
• Contradiction flags. When a new source disagrees with an existing page, the librarian does not pick a winner. It flags the conflict for a human, because deciding truth is governance, not formatting.
• A search index over the wiki, keyword and semantic together, so the answer to most questions is one small, current page instead of an archaeology dig.

# 07:10 every morning: ingest new sources, update pages, open a PR
10 7 * * * cd ~/team-wiki && ./librarian ingest sources/ --open-pr

A wiki rots the moment its readers stop trusting it, and trust dies through small things. Date-stamp claims, so stale advice looks stale. Keep pages short enough that updating one is a five-minute act, not a project. Let the librarian propose and humans approve, the same reads-free, writes-governed rule that protects any shared brain. And measure it honestly: when someone asks a question the wiki cannot answer, that is a missing page, and the librarian should draft it the same day.

# Agents and people both: search the wiki before the sources
wikigrep "payment retry policy"   # keyword, instant
wikiask  "why did we drop provider X?"   # semantic, when you can't name it

Real pieces to build it from

ObsidianLocal-first markdownLinked markdown pages with backlinks and graph view. A wiki that both humans and models read natively.GitHub ActionsScheduled CIRun the librarian on a schedule with the wiki in a repository, so every change is a reviewable pull request.Claude CodeAnthropicHeadless mode makes a capable librarian: read sources, edit pages, open the PR, all from one scheduled command.RIripgrepOpen sourceThe keyword half of retrieval. Over a folder of short markdown pages, it is effectively instant.CHChromaOpen sourceLocal embeddings for the semantic half, so questions find pages even when the words do not match.NotionHosted alternativeIf the team already lives here, the same pattern works through its API; the librarian writes pages instead of files.

Teams that run this for a few months describe the same shift: the wiki stops being documentation and becomes the organisation's working memory. New joiners read it and are dangerous in a week. Agents read it and stop hallucinating your own business back at you. And because the librarian never gets bored, the corner pages that no human would ever maintain stay current too. It is the least glamorous system we build for clients, and very possibly the highest-return one.