Feed it notes, papers, and links. An AI builds and maintains a real markdown wiki on your disk — and asks for your approval before every change. Not RAG: knowledge that compounds.
Three layers on disk, exactly as Karpathy's LLM Wiki gist describes — your raw notes stay yours; the AI owns only the wiki layer.
Quick captures, full notes, PDFs, URLs. The AI reads this layer but never rewrites it.
It digests your sources into context.md and research pages — cross-referenced, deduplicated, contradiction-flagged.
Questions hit the maintained wiki, not raw chunks. Good answers get filed back as new pages. Knowledge compounds.
Every note carries a status chip — ✨ Add to wiki until digested, ✓ In wiki after. The AI proposes takeaways and a checkbox plan first; nothing is written without your approval.
One click and the AI re-reads your whole wiki: contradictions with the exact conflicting sentences quoted, stale claims, orphan pages, open gaps — as cards you can act on. Notion and NotebookLM cannot do this.
No subscription, no API key, no data leaving your machine.
~/mindbase-data — grep it, git it, leave anytime| llama3.2:3b | 8 GB RAM | capture & search |
| qwen3:14b | 24 GB+ | interactive default |
| muse-glimmer:30b | 32 GB+ Apple Silicon | deep synthesis & lint |
npx mindbase-app
Starts the local server, opens the web app, walks you through picking a free local model. Node 20+.
/plugin marketplace add frankchu91/mindbase-llm-wiki /plugin install mb@mindbase
Slash commands (/mb:contribute, /mb:build, /mb:lint…), sub-agents, auto context injection.
{ "mcpServers": { "mindbase": {
"command": "npx",
"args": ["-y", "mindbase-mcp"]
} } }
49 tools: ingest, search, synthesize, health-check your wiki from any MCP-compatible editor.