I built an ICM vault to learn ICM (Claude Code + Obsidian)

I kept watching Jake's videos, reading the ICM paper, Karpathy threads… and forgetting most of it a week later. Learning the wrong layer, but with my own knowledge.

So I built the thing the content kept pointing at. Not a notes app. An ICM workspace for my own learning.

The rules (sound familiar?):

- raw/ is immutable. Transcripts, PDFs, papers. The LLM reads, never edits.

- wiki/ is the LLM's workspace: entities, concepts, sources, analysis, overview.

- index.md = catalog. log.md = append-only timeline. Welcome.md = the schema the agent follows.

- Claude Code writes. Obsidian is the front-end. I read; it writes.

- Every claim links to a source. No orphan facts.

One ingest: I drop a transcript in raw/ and say "ingest it." A single 22-min Van Clief video cascaded into ~10 linked pages in one pass (a summary, 2 new entities, 5 updated pages, index + log), cross-linked, with contradictions flagged rather than silently resolved.

After ~3 months: ~35 sources, 84 interlinked pages, every claim cited. No vectors, no RAG. The folder path is in the memory, and I can read every byte.

I'm using the method to learn the method. Folders over agents, pointed at my own head. (I'm starting an AI-implementation practice in Brazil built on ICM, so this doubles as the knowledge base behind my client work.)

Schema's attached if you want to start your own. What would you ingest first?https://github.com/jarantes-prog/icm-knowledge-vault

6 comments