An LLM wiki instead of RAG

Osiris gives Claude a long-term memory. It does not need a vector database for that, but the pattern Andrej Karpathy described in April 2026 as the “LLM wiki”: after every conversation, a separate run reads what was said and writes it into Markdown pages. The knowledge is compiled when it is written — not pieced together again on every question.

What RAG does — and what it leaves open

Retrieval-augmented generation, RAG for short, cuts documents into chunks, turns them into vectors and, for every question, fetches the chunks that look most similar. For large collections of documents that works. As a memory of your own work it has a catch: every question starts again from raw material. Two chunks that contradict each other both come back, and what was understood in one conversation is not there in the next — the model has to piece it together anew each time.

The LLM wiki: compile once, then look it up

Karpathy's idea turns the order around. The understanding happens up front: the model reads a source once, summarises it and files what it learned under the topics it belongs to — adding to a page that already exists instead of writing a second one next to it. What gets searched later is no longer raw material, but pages that have already been written, linked and checked. In Osiris this is the Memory: the inside of the brain, what it keeps long-term.

Three layers with clear ownership

• raw/ — the sources: every conversation and every file you put there, unchanged. The agent reads them and never rewrites them.
• wiki/ — the pages the agent writes: entities, concepts, one summary per source, comparisons and answers worth keeping. Every paragraph names the source it comes from.
• schema.md — the rules of the Memory: how pages are written and when a page is extended instead of duplicated. A file you can read and change.

Next to them, Osiris keeps two files itself: index.md, the map of all pages with one sentence each, and log.md, one line per operation, append-only. Nothing happens unnoticed.

Ingest, query, maintain

• Ingest — ten minutes after the last message of a conversation, once per conversation. The run writes exactly one summary of the source, extends the pages it touches, and when the new source contradicts a page, it corrects the page and names the contradiction with date and source.
• Query — the map first, then precisely the pages that fit. The answer says which page and which source it rests on; a synthesis that did not exist yet is filed as a new answer page.
• Maintain — once a week by default: contradictions between pages, outdated statements, orphaned pages, missing cross-references and sources without a summary.

Found in English and in German

Every page carries two extra fields in its header: aliases, other names for the same thing — in both languages, with abbreviations and old names — and fragen, the questions the page answers, phrased the way someone asks who does not know the answer yet. That way a page is found even when the question uses different words than the page.

Memory and brain: what happened, what applies

The Memory sits next to the brain, not inside it. The brain holds what applies — rules, notes and decisions that you write down and that Claude follows. The Memory holds what happened — condensed from your conversations. One binds, the other remembers. How Claude gets to both is described under Claude Code with memory.

Plain Markdown on your disk

The Memory is a folder of Markdown files on your machine. You can read the pages, open them in Obsidian or put them under Git. Only what a single request needs goes to Anthropic. And like any memory it can be wrong: what Osiris remembered was right once — the weekly maintenance run looks for contradictions, but it does not find every one.