Karpathy's hand-curated LLM wiki, without the hand-curation
The shape of a personal knowledge base for the AI era, with the AI doing 80% of the upkeep.
The pain
Andrej Karpathy described the problem: keep a personal LLM wiki of the answers worth remembering. Three layers (raw transcripts, distilled wiki pages, structured schema) and three operations (ingest, query, lint). He maintains his by hand because no consumer surface offers the right shape.
That works for Karpathy. For most people, the hand-curation tax is too high. So they don't do it, and the answers leak away.
What Memory.Wiki does
Same shape as Karpathy's. Different effort:
| Layer | His way | Memory.Wiki way |
|---|---|---|
| Raw | Manual transcript copy + folder organization | One-click capture from any AI tool, automatic permanent URL |
| Wiki | Hand-write distilled pages | Auto-synthesis with diff/accept. New captures generate proposed wiki updates, you accept or skip |
| Schema | Hand-tag entities + relationships | Embeddings + semantic graph + cross-ref rollup, all automatic |
| Ingest | Manual | /memory.wiki capture from any coding agent + paste/file drop |
| Query | grep + read | /recall API: vector + BM25 hybrid, paragraph-level chunks |
| Lint | Periodic manual review | Hub lint runs automatically and surfaces gaps, conflicts, orphans |
The architecture is the same. The maintenance burden drops by roughly 80%. You stay in the loop on what enters the wiki layer (the diff/accept UI is intentional friction); everything else is handled.
What this case actually replaces
If you're already running a personal Notion, Obsidian, or DEVONthink for AI outputs and feel the hand-curation tax, Memory.Wiki is the lower-effort version of the same thing. If you're not running anything because the tax was too high, Memory.Wiki is what makes it possible.
See it in shape
- Long-form architecture: How Memory.Wiki works and How Memory.Wiki Memory works.
- Live example hub: Memory.Wiki Foundations bundle.
- Cross-AI verified: MWBench shows 100% accuracy across Claude, OpenAI, and Gemini on truly unseen content.