ICM Knowledge Vault
jarantes-prog/icm-knowledge-vault
Base de conhecimento viva, mantida por IA, estruturada em pastas transparentes
Ver no GitHubUm vault de conhecimento mantido por IA que organiza e compila informações sobre tópicos em movimento rápido usando uma estrutura de pastas legível. Você curador fontes e dirige a análise; o agente (Claude Code) faz o trabalho de bookkeeping — resumos, referências cruzadas, filing e consistência. Funciona com markdown simples em Obsidian ou qualquer editor, sem dependência de embeddings ou RAG.
Instalação
npx convez-ai jarantes-prog/icm-knowledge-vault1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Compilar e manter repositórios de conhecimento sobre tópicos dinâmicos (pesquisa, mercados, tecnologia)
- Criar wikis internos de empresas com sources imutáveis e análise automática
- Organizar aprendizado composto com rastreamento de mudanças e logs de ingestão
Veredito

Padrão elegante que inverte o fluxo convencional de PKM: em vez de você buscar e resumir, um agente lê suas fontes e mantém o vault coeso. Excelente para domínios de movimento rápido onde síntese humana é valiosa mas consistência é crítica — o custo é disciplina na curadoria de fontes e estrutura de schema. Replantável e agnóstico de ferramenta.
README
A tradução para o português deste README ainda está na fila.
ICM Knowledge Vault
An LLM-maintained knowledge base. Folders over agents.
A starter template for a compounding knowledge vault that an AI agent (e.g. Claude Code) reads and writes for you. You curate the sources and direct the analysis; the agent does the bookkeeping — summarizing, cross-referencing, filing, and keeping every page consistent and cited.
Inspired by Jake Van Clief's Interpretable Context Methodology (ICM) ("folders over agents") and Andrej Karpathy's LLM-wiki idea. Nothing here is AI-specific — point it at any fast-moving topic: law, biotech, a market, a company's internal knowledge, a hobby.
You read; the LLM writes. Obsidian is the IDE, the LLM is the programmer, this vault is the codebase.
Why this instead of a notes app or a vector "second brain"
- Glass-box, not black-box. Every fact lives in a plain markdown file you can open, read, and edit. When the agent gets something wrong, you see exactly which file and fix it.
- Position-addressed memory. The folder path is the index. No embeddings, no RAG, no retrieval step — and it stays legible to a human.
- Portable + no lock-in. It's a folder. Clone it, zip it, commit it. Any editor opens it; any capable agent can run it.
- It compounds. One source ingest cascades across many linked pages. The value grows with every add.
How it's structured
.
├── Welcome.md ← the schema the agent follows (read this first)
├── CLAUDE.md ← pointer so Claude Code auto-loads Welcome.md
├── index.md ← the catalog, updated on every ingest
├── log.md ← append-only timeline of what changed
├── raw/ ← IMMUTABLE source documents (the agent reads, never edits)
└── wiki/ ← the agent's workspace
├── overview/ ← top-level syntheses & theses
├── entities/ ← people, places, orgs, products
├── concepts/ ← topics, themes, ideas, frameworks
├── sources/ ← one summary page per ingested source
└── analysis/ ← comparisons, deep dives, filed-back answers
The full rules (frontmatter, naming, ingest/query/lint workflows, citation discipline) live in Welcome.md.
Quickstart
- Click “Use this template” → Create a new repository (or just clone this repo).
- Open the folder in Claude Code (or any agent that can read files). It auto-loads
CLAUDE.md, which points atWelcome.md. - Drop a source into
raw/— a PDF, an article, a pasted transcript, your notes. - Say “ingest it.” The agent writes a cited source summary, creates/updates entity and concept pages, and updates
index.md+log.md. - Later, ask “what do we know about X?” It reads the index, follows the links, and answers with citations.
That's the whole loop: drop → "ingest it" → ask.
Do I need Obsidian?
No. Obsidian is optional — a nice viewer for the graph and [[backlinks]]. The vault is plain markdown in folders, so any text editor (or GitHub) opens it. If you do use Obsidian, add -path:raw in Graph View → Filters to hide the immutable raw sources and see only the linked wiki.
Credits
- ICM / "folders over agents" — Jake Van Clief (Clief Notes community, ICM paper).
- LLM-compiled wiki pattern — Andrej Karpathy.
License
MIT — see LICENSE. Use it, fork it, make it yours.