Your notes, but the AI actually writes them

What it does

Drop files into raw/ and tell your agent to ingest. The agent reads them, extracts entities and concepts, and builds a persistent markdown wiki with auto-generated pages, [[wikilinks]], and a living overview that gets rewritten every time you add a source. It also lints for orphans, flags contradictions at ingest time, and can answer queries by synthesizing across everything you’ve fed it.

The interesting bit

This is essentially a prompt-and-convention layer on top of Claude Code, Codex, Gemini CLI, or OpenCode—not a standalone service. The heavy lifting is the agent reading schema files (CLAUDE.md, AGENTS.md, GEMINI.md) that define wiki structure, then doing the drudge work of cross-referencing and updating. The knowledge graph layers deterministic wikilink edges with LLM-inferred “semantic” edges, complete with confidence scores and ambiguity tags.

Key highlights

• Compiles knowledge once, keeps it current—contrasts itself with RAG’s re-derive-every-query approach
• Auto-converts 15+ formats (PDF, DOCX, PPTX, XLSX, EPUB, even audio) via markitdown at ingest time
• Outputs plain markdown designed for Obsidian; symlink the wiki/ folder into your vault
• Self-contained graph.html visualization—no server, opens in a browser
• SHA256-cached graph builds mean only changed pages get reprocessed

Caveats

• Requires an existing coding agent (Claude Code, Codex, Gemini CLI, etc.)—this is configuration and convention, not a standalone tool
• The “no API key needed” claim is technically true because the agent brings its own credentials; you’re still paying for agent usage
• Multi-format ingest depends on optional Python dependencies (markitdown, arxiv2md, marker-pdf) that need separate installation

Verdict

Researchers, competitive analysts, or anyone drowning in unread PDFs who already lives in Claude Code or Codex should try this. If you don’t use a supported coding agent, or you want a fully packaged app with its own UI, this won’t fit.