Scribe

Scribe is a single binary CLI tool that watches your development work and turns it into a personal knowledge base. It pulls from git repositories, Claude Code sessions, Codex CLI transcripts, and self sent URLs, then organizes what it finds into curated markdown articles you can query from the terminal or from inside an AI agent. The idea is that you stop explaining the same context over and over because the system has already absorbed it and can surface it when relevant.

The problem it tackles is familiar to anyone who uses AI coding assistants heavily. You solve a tricky architecture problem in one project, write something clever in a Claude session, but three weeks later that knowledge is buried in a transcript you will never find. Scribe extracts those decisions, patterns, and learnings automatically, stores them in a local wiki indexed for semantic retrieval, and keeps everything current as you keep working.

Processing happens in two stages. First a keyword density filter runs over raw sources to avoid sending low signal content to the language model. Then a two pass absorption routine expands dense sessions into multiple wiki articles, each tied to a specific entity or decision. The result is a knowledge graph with ten relationship types (supersedes, contradicts, depends on, and so on) that lets you trace how your understanding has evolved. Weekly consolidation cycles detect contradictory claims and resolve them so the base stays coherent.

On the retrieval side, Scribe uses BM25 full text search through a companion tool called qmd. You can run queries from the terminal or inject context blocks into Claude Code and Codex sessions via a handshake protocol, which means the agent consults your knowledge base before suggesting a library you already rejected or proposing an architecture you already tried and documented.

The whole thing runs locally by default. You can point it at Ollama with models like qwen3 or gemma and pay nothing for inference. If you prefer hosted providers, it supports Anthropic Claude, OpenAI compatible endpoints like Together or Groq, and Hugging Face. Scheduling is handled through LaunchAgents on macOS or crontab on Linux, running hourly commits, regular extraction sweeps, and weekly memory consolidation without manual intervention.

The tool is open source under MIT and distributed as a single downloadable binary. It fits developers who code with AI tools frequently and want to stop losing context between sessions, teams that want a shared but version controlled knowledge base, and anyone who values owning their data in a plain markdown format they can read and edit directly.