Most second brains dump everything into one pile. This one separates things into three — Project, Self, and Knowledge — because they have opposite needs. Local, private, in Portuguese, runs on Linux.
# you drop in the raw material │ ▼ 0-inbox/ # captures without classifying │ /triagem # routes to the right brain ┌────┼─────────────┐ ▼ ▼ ▼ 1-projeto 2-self 3-conhecimento o que quem eu fatos / o que rolou sou/penso eu aprendi └────┴─────────────┘ │ Claude reads the vault ▼ IA que já sabe quem você é antes de você digitar.
There are 3 types of memory with opposite needs — corresponding to the 3 types in cognitive science. Mixing them bloats the brain. Each gets its own folder, rule, and skill; the [[wikilinks]] cross-link with one another.
What happened: decisions, state, next actions. Memory episodic — changes every day and is then archived. Append-only and dated. The LLM here records faithfully.
Who you are: values, beliefs, thoughts, open questions. Memory of identity — changes slowly, it’s your. The LLM here questions you, doesn't agree.
Facts about the world, what you learned. Memory of reference — only accumulates; it’s searchable. The LLM here distills raw material into a wiki connected.
It’s not software with a hidden database—the memory is your file system. Claude Code reads the CLAUDE.md at the start of every session; it routes you to the right folder without searching through everything. Since it’s just text, it survives the hype: the model changes, the brain stays.
O CLAUDE.md stores the vault map + the rules (voice, design, where to look). It’s the most important piece.
The LLM reads the raw material and assembles the wiki with [[wikilinks]]. You drop things in and review — you don't curate by hand. The wiki grows and improves every future answer.
Organization (this project) + Format (kepano) + Retrieval (graphify) + Visualization (brain-map). Only the 1st is required.
The core only needs git + Claude Code. The layers and file ingestion are optional.
The agent that reads the vault. (Codex works too.)
# install Claude Code https://claude.ai/code
To clone and version the brain.
# Linux sudo apt install git
To ingest PDFs, spreadsheets, and docs into brain 3.
# key in .env GEMINI_API_KEY=...
Real commands. The setup.sh creates the vault for the 3 brains; you do everything else by talking with Claude.
Creates the vault with the structure for the 3 brains (by default in ~/cerebro).
git clone https://github.com/inematds/cerebro-inema.git cd cerebro-inema ./setup.sh # or: ./setup.sh ~/meu-cerebro
Open Claude Code in the vault folder and run setup. It interviews you in plain text and fills in the CLAUDE.md and the Self brain.
cd ~/cerebro # in Claude Code: /setup-cerebro
Throw anything into 0-inbox/ without thinking about where. Then route it to the right brain.
# drop notes, links, PDFs in 0-inbox/ — then: /triagem # classifies and moves to the right brain
Turn a raw source into a connected atomic note (brain 3), with valid wikilinks.
/nota # source → wiki note in 3-conhecimento/notes/
When you want tension instead of validation: the LLM challenges you and raises contradictions with your values.
/socratica # the LLM questions, it doesn't agree
Saves decisions, lessons learned, and next actions to the right brain + updates the memory.
/daily # starts the day with context /tldr # saves the session summary in the right place
No need to open Obsidian: the brain is the files, and Claude works with them.
# you: joguei 5 coisas na inbox, organiza # /triage proposes: proposta-cliente.pdf → 1-projeto (decisão/cliente) "e se eu cobrasse por valor?" → 2-self (questions/) artigo-pricing.md → 3-conhecimento (/nota) ideia-canal-yt.md → 2-self (thinking/) nota-do-dia.md → 1-projeto (daily/)
# you: acho que devo abrir um curso novo # /socratica: seu values.md diz "foco antes de escala". um curso novo contradiz isso agora? o que te faria não fazer? quem discordaria — e por que estaria certo?
Only organization is required. The other three are optional foundations—install them as the brain grows.