MindBase: AI wiki that learns from your notes, now in beta

MindBase is a local-first AI research assistant that maintains a personal markdown wiki from your sources. The product implements Andrej Karpathy's LLM-Wiki pattern and entered early access with a beta feedback period scheduled for 2026-Q3. It works by organizing three physical layers on disk: Layer 1 is an append-only sources/ directory containing contributors/ (your dated thoughts, like a lab notebook), research/ (wiki pages the LLM writes), and raw/ (original PDFs, URL captures, pastes). Layer 2 includes a README.md file (the rules you set for your wiki), context.md (the synthesized current truth the LLM writes), and index.yaml (an auto-generated file catalog). Layer 3 holds logs/ (a chronological record of every operation) and artifacts/ (generated outputs like daily briefs, exports, lint reports).

Three core operations form a loop. Contribute: you feed a thought, PDF, or URL; the LLM reads it, discusses key takeaways with you, updates 5–15 relevant wiki pages, and appends a log entry. Ask: you pose a question; the LLM reads your already-synthesized context.md first, then drills into cited source pages and returns answers with [[wikilink]] provenance. Lint: the LLM audits the whole wiki for contradictions, stale claims, orphan pages, missing cross-references, and gaps that need investigation. Good answers and lint findings flow back in as new contributions, so the wiki compounds over time. This stateful synthesis model differs fundamentally from RAG tools like NotebookLM or Perplexity, which re-derive answers from raw sources every query with nothing accumulating, and from passive containers like Notion or Obsidian, where the human organizes and AI features are add-ons.

MindBase ships as an MCP (Model Context Protocol) server published on npm as mindbase-mcp. It runs entirely on your machine with no cloud upload, signup, or telemetry. Claude Code receives the flagship experience: users install via two commands (/plugin marketplace add frankchu91/mindbase and /plugin install mb@mindbase), gaining access to 12 slash commands (/mb:contribute, /mb:ask, /mb:build, /mb:status, and 8 others), 5 sub-agents with security boundaries (contributor, builder, curator, researcher, migrator), and a SessionStart hook that auto-injects your project's README, context, and recent log entries into every new session. The Claude Code plugin also implements the full Karpathy 8-step ingest: read source fully, discuss 3 key takeaways with user, propose which wiki pages to create or update, wait for approval, then execute and confirm. Other MCP-compatible editors (Cursor, Windsurf, Cline, Continue.dev, Zed, Aider, Goose, and future custom SDK apps) access the same core MCP tools via a one-line config entry pointing to mindbase-mcp as a stdio server, but without the Claude Code plugin's extra UX layer. Installation requires Node.js 20+ and pnpm 10+ (only for Claude Code plugin or web UI from source); no clone or build is needed because the MCP server installs itself via npx. The project is open-source (github.com/frankchu91/mindbase) and the team noted that the core data model and loop are stable, though some UI features are still migrating from v1 to v2.

MindBase addresses a real gap in the knowledge-management landscape. Existing tools force users into one of two camps: either passive containers (Notion, Obsidian, Roam) where the human does all the organizing and the AI is bolted-on generation; or stateless RAG systems (NotebookLM, Perplexity, ChatGPT search) that re-derive answers from raw sources on every query, meaning no knowledge accumulates. By implementing Karpathy's LLM-Wiki pattern as a working product, MindBase proposes a third option where the LLM is a maintainer that actively writes and updates synthesized wiki pages as sources flow in. The design is stateful: each contribution triggers the LLM to read new material, cross-reference it with existing knowledge, flag contradictions, and update relevant wiki pages. Over time, the synthesized context.md file becomes the source of truth, richer and more consistent than raw sources alone.

The architecture is deliberately transparent and local-first: three physical layers on disk (sources, markdown wiki files, logs) with no proprietary database or cloud dependency. This appeals to users who want to own their data and understand their system's data flow. The product also distributes as an MCP server, plugging into the ecosystem of AI editors rather than trying to build its own IDE — Claude Code gets a flagship experience with structured sub-agents and slash commands, while other MCP-compatible editors (Cursor, Windsurf, Cline, Continue.dev) can access the same core tools with lighter UX. The team framed MindBase explicitly as early access with core data model and loop stable but some UI features still in migration, signaling they're testing the idea in the wild before consolidating the experience.

No comments yet. Be the first to share your thoughts!