Setup Knowledge OS
tillmannvey-spec/setup-knowledge-os
Sistema de memória persistente e autossustentável para Claude Code
Ver no GitHubSkill interativo para Claude Code que configura um sistema de conhecimento autossustentável, onde cada sessão de codificação automaticamente alimenta a próxima com contexto persistente. Implementa memória em camadas, Vault em Markdown, Switchboard (indexador local) e hooks de integração, com foco em isolamento e segurança dos dados.
Instalação
npx convez-ai tillmannvey-spec/setup-knowledge-os1 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
- Configurar um sistema de memória que persiste conhecimento entre sessões de codificação
- Migrar ou reconstruir uma base de conhecimento existente em estrutura organizada
- Manter índice automático de projetos e conceitos com sumarização por IA
Veredito

Ferramenta sólida e bem arquitetada para quem trabalha com múltiplos projetos simultâneos no Claude Code e precisa de memória real entre sessões. A abordagem safety-first (backups, isolamento, diffs) é rara e prudente; exige dedicação para configuração inicial, mas o payoff é genuíno.
README
A tradução para o português deste README ainda está na fila.
setup-knowledge-os
An interactive onboarding skill for Claude Code that walks you, step by step, through setting up — or rebuilding/migrating — a self-maintaining knowledge OS: a system where every coding session automatically becomes context for the next one, so your tool stops forgetting between sessions.
This is the skill behind the CTA of the explainer video "My tool doesn't forget anymore."
What's a skill? A folder with a
SKILL.mdthat Claude Code loads on demand. Drop this folder into~/.claude/skills/, and Claude can invoke it when you ask to set up the system. See the Claude Code skills docs.
What the system is
The knowledge OS this skill installs has four pillars plus the hooks that wire them together:
- Memory system — four layers, each with one clear job (per-project auto-memory, a global Vault index, project
CLAUDE.mdfiles, and global rules). - Vault — human-readable Markdown (entities / concepts / summaries) that the agent maintains and cross-links itself. No black-box database.
- Switchboard (Agentic OS) — a local Node service that indexes your Claude-Code sessions into SQLite, has a cheap model write one-sentence summaries, and patches them back into your project frontmatter. The "busy worker" that keeps the Vault current.
- Lookup protocol — find facts (live URLs, branches, status) by rule instead of guessing or grepping.
The loop: you work → a Stop-hook drops a marker → the Switchboard condenses it → a memory index regenerates → a SessionStart-hook reads it back into context. The circle closes. You type your first word and Claude already knows where you left off.
One folder, every session
A core design choice: everything lives in one root folder. You open that one folder in a single VS Code session — and every Claude-Code session then has access to all your projects and all your knowledge at once, with the agent able to cross-reference across them. The repo's references/folder-structure.md explains the bucket layout (_os/ clients/ apps/ concepts/ products/ experiments/ + INDEX.md / DEPLOYMENTS.md) and — importantly — the skill asks you whether to adopt this structure or design your own.
What this skill guarantees (safety-first)
The setup it performs is potentially invasive (hooks, scheduled tasks, moving paths). So the skill is built around three iron rules:
- Isolation first — it works in a git worktree (for tracked targets) or timestamped backups (for things like
settings.json), and only applies to your live tree after you approve the diff. - Never delete — nothing is removed; anything retired is moved to
_archive/. Norm -rf, nogit reset --hardon your files. - Paths get updated everywhere — when a path, folder name, or port changes, it finds every reference (in all encodings — double-spaces,
\vs/, JSON-escaped, base64url,%20), shows you each one, updates them individually, and re-greps to prove nothing was left dangling.
Install
git clone https://github.com/tillmannvey-spec/setup-knowledge-os.git \
~/.claude/skills/setup-knowledge-os
Then in Claude Code, just say what you want — e.g. "set up the knowledge OS" or /setup-knowledge-os — and the skill takes over.
Contents
| File | What it is |
|---|---|
SKILL.md | The interactive 8-phase walkthrough + the three safety rules |
references/system-components.md | Component map + every place paths are hardcoded |
references/path-migration.md | The "update every reference, everywhere" protocol |
references/folder-structure.md | The mono-root philosophy + adopt-or-customize decision |
Note
The skill's component map describes a concrete reference implementation (Windows / PowerShell, a Node Switchboard on ports 4123/4242). The walkthrough is honest about flagging what needs adapting on other systems rather than blindly proceeding. The Switchboard service itself is a separate project; this repo is the onboarding skill only.
License
MIT — see LICENSE.