SecondBrain

The idea comes from the “LLM Wiki” concept (Karpathy, 2025): a private Obsidian vault where Claude Code acts as the editor. I supply the material — articles, documents, recordings, thoughts — and the agent turns it into linked pages, keeps things in order and suggests the next questions.

The architecture has three layers. Sources are raw material, untouched once processed. The wiki is processed knowledge: entities, concepts, decisions, guides and maps of content, tied together by a thousand links. The schema is the rulebook the agent works by — and it evolves together with the system.

Knowledge should be trustworthy, not just plentiful. Every claim leads back to a source, contradictions are flagged explicitly, uncertainty is spelled out (“per source X”, “as of…”), and each page records its provenance: extracted or inferred. A new page is created only when a topic shows up in at least two independent contexts — a barrier against zombie notes. A regular audit catches dead links, orphans and drift.

The vault works from anywhere, not just inside Obsidian. Through an MCP server and global skills I can, from any project, save an insight or a decision, or ask about something I already know. Meeting recordings are transcribed locally (WhisperX) and turned into a note — markdown plus a ready document. Everything lives in git, and an off-machine backup runs automatically at session start.

The system improves itself. When I do something by hand a second or third time, the agent proposes a skill for it. When a rule starts to chafe, it proposes a concrete schema change. The decision is always mine. RailwaySecondBrain (a citable base of railway regulations) and the notes search in KomPi are built on this foundation.

The graph alongside is the real link map of the vault, generated locally and anonymised: no titles, no content, just the shape of the knowledge. It shows how AI, projects and railway work interweave.