LLM Wiki: Build a Knowledge Base Your AI Maintains
Set up an LLM wiki in a folder of Markdown files. Follow one team decision from source memo to cited answer, then update it when the decision changes.
Published

The next time someone asks why your team still sends launch invitations by hand, you could give them a page with the decision, its reason and the condition for changing it. An LLM wiki lets an AI maintain that page between sessions. You choose the evidence and decide when the answer has changed.
What survives the chat
In his April 4, 2026 note, Andrej Karpathy describes a language model that “incrementally builds and maintains a persistent wiki.” The part worth keeping is the connection someone already worked out. A release memo might say what the team chose. A later engineering note might explain which constraint has gone away. A wiki page can put both in order and link to the passage behind each claim. Karpathy’s original LLM Wiki gist
Think of a handbook editor with a filing cabinet. The editor can revise the handbook every time a signed memo arrives. The memo stays in the drawer, available to anyone who wants to check the wording. Karpathy gives these parts specific names:

Set up a wiki for one release decision
Choose a coding agent that can read and edit a folder. Open that folder in your usual text editor, or use Obsidian if you like browsing linked notes. The filenames below are a starter layout. The sample memos are fictional so you can try the whole workflow without exposing company records.
Make a folder named release-wiki/. Inside it, create raw/ for originals and wiki/ for pages. Put two empty files in the latter: wiki/index.md for a linked catalog with a sentence about each page, and wiki/log.md for a chronological record you only append to. Put SCHEMA.md at the top of the folder. Tell the agent to read it at the start of each session. If your agent loads a project instruction file automatically, you can keep the rules there. How a team instruction file works
Give the agent a filing rule it can actually follow. Paste this into your schema:
Read originals in raw/ and leave them intact. Write pages under wiki/. On each decision page, record the current rule, its reason, what would change it and links to the dated originals. Mark missing information as unknown. On ingest, update affected pages and the index. On query, start from the index and cite the supporting passage. On lint, list conflicts and broken links for review. Append each operation to the log. If sources disagree, show both and ask the decision owner.
Ingest the first memo. Create raw/pilot-decision.md with this sample text: “Pilot kickoff, approved by the release lead. Send invitations by hand because account recovery is untested. Revisit after the recovery drill.” Ask the agent: “Read SCHEMA.md and ingest raw/pilot-decision.md. Show me the changed files.” Look for a linked decision page, an index entry and a log entry. Open the source link. The page should say the drill triggers a review. Passing the drill does not itself authorize automatic invitations.
Query it from a fresh session. Open the same folder and ask: “Read SCHEMA.md. Can we automate invitations yet? Who can approve the change? Cite the source.” A grounded answer should give the manual rule and its reason. The kickoff names the person who approved that memo, but leaves authority for a future change unspecified. If this distinction keeps coming up, ask the agent to file the checked answer and link it from the decision page.
Lint after a change. Add raw/release-amendment.md with: “After the recovery drill, the release lead approved automatic invitations for staff accounts only. External pilot accounts remain manual.” Ingest it. Then ask: “Lint every wiki page that mentions invitations. Show the audience and source for each rule, flag disagreements, and list pages missing from the index.” Check the saved answer as well as the decision page. A new session should now distinguish staff from external accounts and point you to the amendment.

Which work pays back the upkeep?
Start with a question that interrupts someone who has already answered it. Today that person finds the thread, reconstructs the sequence and writes another reply. With a wiki, the decision owner checks the page when a source changes. Subsequent readers can follow the link themselves. To estimate hours saved, add up time spent answering that question in a typical week and subtract the time spent checking edits. A question nobody repeats is a poor place to start.
When to use a wiki and when to use RAG
Choose the wiki when several people keep asking for the reasoning behind the same decisions and an owner can review changes. Retrieval-augmented generation (RAG) looks up relevant passages when the question arrives and gives them to a model to compose an answer. That makes it a better starting point for a large collection full of one-off questions. A detail that seemed too minor for the wiki editor may be exactly what a reader needs. RAG and fine-tuning for company knowledge
Freshness depends on the update path. For “Who is on call right now?” read the current rota directly, or retrieve from an index whose refresh delay the team accepts. A perfectly cited wiki page can still describe yesterday’s shift. Keep the explanation of how to escalate in the wiki and link out to the rota for the name.
Many writers change the choice. If separate teams edit originals all day under different access rules, copying their conclusions into a shared folder creates a second publication process. Prefer retrieval that enforces the original permissions and picks up revisions. If several people do maintain a wiki together, give each decision an owner and queue proposed edits for review before merging them.
Keep secrets and customer data outside the folder. Exclude credentials, tokens, customer messages, account records and identifiable call transcripts before the agent reads anything. A summary can repeat a private detail onto a broadly shared page. For support work, write an approved, anonymous description of the product behavior and ingest that instead.
Fix the two ways a good page goes bad
A developer in the gist discussion describes a file reference going stale as soon as something is renamed. A lint pass can check whether the target exists. A claim needs another check: open the replacement document, confirm it still supports the sentence, then update every page that cites the old reference. If the evidence is gone, mark the claim unverified. Practitioner’s account of stale references
Another builder describes tracking which decisions supersede others to address stale and contradictory notes. Keep both sources, record their effective dates and ask whether the apparent conflict is really a difference in scope. In the invitation example, both rules apply after the drill, each to a different audience. When two approved memos really do disagree about the same people and period, put the question in front of their owner. Record the resolution before offering a definitive answer. Builder’s account of tracking replaced decisions
Two products a small team could build
Build the decision-change reviewer first. Sell it to an operations lead who owns a handbook but has to chase subject experts for corrections. The smallest useful version accepts an approved memo, proposes edits to affected pages, and shows each changed sentence beside its source for the owner to accept or reject. The difficult part is finding the footnote on another team’s page that quietly repeats the old rule. The specific pitch to validate with them is whether reviewing changes costs enough time to justify a dedicated tool.
Build a role-specific onboarding checklist. A people-operations lead could choose a role and generate a first-week checklist from approved handbook pages, with a source and owner beside each task. Start with one role and show which tasks a handbook revision would change. The buyer worth interviewing hires into the same role repeatedly and has already had to correct an outdated checklist.
Questions teams ask
How can I create an effective internal knowledge base?
Start with a decision that people regularly ask someone to explain. File its approved source and make a page that answers the question. Hand it to a teammate who missed the decision and see whether they can find the reason and the evidence.
What is the best knowledge management tool?
For this pilot, use the folder editor your team already knows. If people need separate access rights, approvals or managed search, compare those features against the way you publish decisions today. Knowledge management tools compared
What are the steps involved in the onboarding process for an employee?
For a role-specific checklist, group tasks by what the hire needs next: arrange access before arrival, introduce owners on day one, guide the first piece of work, then schedule a check-in. Have the role owner approve the sequence. Keep personal progress and account details in your existing onboarding system.
If you want a source-linked review workflow built for your team, see AI automation services.
- Published
- Category
- Build







