The Problem with How We Store Knowledge
- Explain why traditional notes apps create write-only memory and fail as long-term knowledge bases
- Describe the core limitation of RAG systems and why re-deriving knowledge on every query prevents accumulation
- Articulate the Karpathy insight: the difference between compiled persistent knowledge and retrieved knowledge
- Identify at least three use cases where the LLM Wiki pattern creates compounding value over time
The Notes App Trap
Most people have tried some version of the same system: a notes app, a folder of bookmarks, a Notion database. You save things religiously — articles, ideas, research, links you mean to revisit. And then almost never find them again, not because the tools are bad, but because saving is only half the problem. The other half is synthesis, and no notes app does that for you.
Researchers call this "write-only memory." You add things in. Nothing useful comes back out. The knowledge sits there accumulating but not compounding. You are not getting smarter. You are hoarding.
Why RAG Falls Short
When AI tools entered the picture, they brought a pattern called RAG — Retrieval-Augmented Generation. The idea: upload your documents, ask a question, and the AI retrieves relevant chunks and generates an answer from them.
This is better than a notes app. But it has a fundamental limitation: the AI is rediscovering knowledge from scratch on every single question. There is no accumulation. Ask something that requires synthesizing five documents and the AI has to find and piece together the relevant fragments every time. Nothing is built up. NotebookLM, ChatGPT file uploads, and most RAG systems work this way.
The Core Insight: Compiled vs. Retrieved Knowledge
In April 2026, AI researcher Andrej Karpathy published a document describing a different approach. Instead of retrieving from raw documents at query time, he proposed having the LLM incrementally build and maintain a persistent wiki — a structured, interlinked collection of markdown files that sits between you and the raw sources.
The key difference: knowledge is compiled once and kept current, not re-derived on every query. When you add a new source, the LLM does not just index it. It reads it, extracts the key information, and integrates it into the existing wiki — updating relevant pages, noting where new data contradicts old claims, strengthening the evolving synthesis.
The cross-references are already there. The contradictions have already been flagged. The synthesis already reflects everything you've added. Every source you add makes the whole richer.
What This Looks Like in Practice
The pattern works across a surprisingly wide range of use cases:
- Personal knowledge: Track your own goals, health, ideas, and self-improvement. File journal entries, articles, podcast notes. Build a structured picture of what you know and think over time.
- Research: Go deep on a topic over weeks or months. Read papers, articles, and reports and build a comprehensive wiki with an evolving thesis — the synthesis gets richer with every source you add.
- Reading a book: File each chapter as you go, building out pages for characters, themes, and ideas. By the end you have a rich companion wiki, like a personal fan wiki for anything you read.
- Business and team: An internal wiki fed by Slack threads, meeting transcripts, project documents, and customer calls. The wiki stays current because the LLM does the maintenance no one on the team wants to do.
The Idea Has Deep Roots
Karpathy credits Vannevar Bush's 1945 essay "As We May Think," which described a hypothetical device called the Memex — a personal, curated knowledge store with associative trails between documents. Bush's vision was closer to this than to what the web became: private, actively curated, with the connections between documents as valuable as the documents themselves. The part he could not solve was who does the maintenance. That is the part the LLM now handles.
Why Wikis Die (and Why This One Won't)
Every team has tried to build a shared wiki. Most are graveyards within a year. The reason is not that people stopped caring — it is that the maintenance burden grows faster than the value. Updating cross-references, keeping summaries current, noting when new data contradicts old claims, maintaining consistency across dozens of pages: humans abandon wikis because this work is tedious and relentless.
LLMs do not get bored, do not forget to update a cross-reference, and can touch fifteen files in a single pass. The wiki stays maintained because the cost of maintenance is near zero. Your job is to curate sources and ask good questions. Everything else is the LLM's job.
- Notes apps solve the capture problem but not the synthesis problem — you save things and then never find or use them in a connected way
- RAG systems retrieve from raw sources on every query; an LLM Wiki compiles knowledge once and keeps it current so every query builds on prior synthesis
- The maintenance burden — updating cross-references, resolving contradictions, keeping summaries current — is what kills human-maintained wikis. LLMs handle this at near-zero cost
- Karpathy's pattern is descended from Vannevar Bush's 1945 Memex concept: private, curated, with connections between documents as valuable as the documents themselves