AI Knowledge Management · Tech Brief

LLM Wiki — When AI Writes Its Own Wiki,
Knowledge Compounds

Beyond RAG's start-from-scratch retrieval: a knowledge-management paradigm where AI writes and maintains its own wiki — what improves, and when to choose which.

Jihwan Woo · May 2026

← Back to Home

📄 Full Report (PDF)

"I Asked Yesterday — Why Is It Searching from Scratch Again?"

Ever uploaded documents to ChatGPT or Claude and asked questions? At first it feels magical. But after a few days comes a subtly frustrating moment: yesterday the AI synthesized five sources into a great answer, and today, on a similar question, it is digging through the sources from scratch again. Yesterday's conclusion survives nowhere.

Most AI document workflows today use RAG (Retrieval-Augmented Generation). Picture a library: every time you ask, a librarian pulls books off the shelves, summarizes an answer, reshelves the books — and takes no notes. No matter how much you ask, knowledge never accumulates. You are earning simple interest, never compound.

The New Approach: AI Writes the Wiki Itself

Proposed by Andrej Karpathy — OpenAI co-founder and former Tesla AI lead — the LLM Wiki flips the problem: instead of reassembling sources on every question, have the AI build a curated wiki and keep updating it.

Back to the library: this time, whenever the librarian pulls a book, she writes key points in a notebook. New books update existing notes, contradictions get flagged, related entries get linked. Over time, the notebook becomes a map of the entire library.

Three-Layer Structure
Raw Sources → Wiki (LLM-written Markdown) → Schema (operating rules)
Ingest

New material updates related pages together

The LLM reads, summarizes, and updates related pages — a single source can touch 10–15 pages at once.

Query

Exploration itself becomes knowledge

Answers draw on the wiki, and good answers are saved back as new wiki pages.

Lint

Periodic cleanup of contradictions and broken links

Regular checks for inter-page contradictions, orphan pages, and missing cross-references.

Why Humans Couldn't Do This Themselves

Wikis are not a new invention — yet personal and small-team wikis almost always fail. The reason is simple: maintenance is too tedious. Every new source means touching 5–10 related pages, correcting outdated information, fixing broken links. After a month or two, people give up, and the wiki rots in a half-current state.

LLMs are the exact opposite: they never get bored, never forget to update cross-references, and can edit 15 files at once. With maintenance costs far below human levels, the wiki gains value as it grows — the main reason humans abandoned wikis is neutralized. Karpathy reported running a wiki of ~100 sources and hundreds of pages with nothing but Markdown files and a single index. Note this is an n=1 self-report; generalization to organizational settings requires separate validation.

How It Differs from Naive RAG

DimensionNaive RAGLLM Wiki
Knowledge accumulationFrom scratch every timeKeeps compounding
Repeated questionsRe-derives the answerAnswers from curated conclusions
Contradictory informationGoes unnoticedFlagged by the AI
Human burdenSources, questions, review — all of itCuration and review only
Integrating new sourcesSimple indexingAuto-updates related pages

This complements rather than replaces RAG. Precision search over millions of documents, real-time data, and access control remain RAG territory. In practice, the realistic answer is a hybrid: the wiki holds "curated conclusions" while RAG handles "precision search over originals."

Who Uses It: Five Representative Scenarios

01

PhD research wiki

After 100+ papers, you no longer remember what you saw where. When every new paper generates a summary page auto-linked to authors, methods, and results, citations and comparisons are instantly retrievable at submission time.

02

Startup technology decision wiki

For teams whose decision rationales are scattered across Slack and Notion — a living record where decisions and their reasoning accumulate.

03

Consulting firm's per-client knowledge base

Client context that used to live only in a departing consultant's head becomes an organization-readable synthesis.

04

Engineering team's incident retrospective wiki

Postmortems stop sleeping as isolated documents and grow into a cross-linked knowledge web of similar incidents, root causes, and preventions.

05

Medical/legal case database

In domains where judgment sharpens as cases accumulate, curated case comparisons become instantly retrievable assets.

A Different Layer from Agentic AI

If agents are "hands that do work for you," the LLM Wiki is the "organizational memory" those hands consult. Where agent harnesses address execution reliability, the LLM Wiki ensures that each task's results and context are reused — compounding — in the next task. When the two layers combine — agents do the work, results accumulate in the wiki, and the next agent starts by reading it — an organization's AI usage shifts from simple to compound interest.

Conclusion: Not "Which Is Better" but "When to Use Which"

The LLM Wiki is not a RAG replacement; it fills a layer that has been empty — the accumulation of curated knowledge. If your sources number in the dozens to hundreds, you ask repeatedly about the same topics, and reusing conclusions matters, the wiki is your answer. If you need precision search over millions of documents with real-time freshness, that is RAG. And most organizations will end up using both.

LLM Wiki Knowledge Management RAG Karpathy Agentic AI Tech Brief

Related Articles

Harness Engineering — Systems, Not Models, Make Agents Kimi K3 — What Happens When Frontier AI Goes Open

Need a talk, article, or advisory on this topic?

I deliver executive keynotes, contribute industry reports, and advise on AI strategy.

Get in Touch