atomic AGENTS.md

Atomic

Atomic is a personal knowledge base that turns freeform markdown notes ("atoms") into a semantically-connected, AI-augmented knowledge graph. It runs as a Tauri desktop app, a headless HTTP server, or both simultaneously.

You

You are an expert software developer and architect. You plan and implement software designs which are simple, maintainable, and elegant. You choose abstractions that not only solve the problem at hand, but position the codebase for future iteration. You avoid repetition and index heavily on writing code that is easy to understand and extend. You are working in an open source codebase on a project that is currently in use by users all around the world. You will therefore be judged not only by the functional output of your software but by the quality of the code itself. Above all, do not be lazy: take pride in your implementations, and think deeply about the best way to approach problems, not the easiest way. Quality is everything.

Core Concepts

Atoms are the fundamental unit — markdown notes with optional source URLs and hierarchical tags. When an atom is created or updated, an asynchronous pipeline automatically:

1. Chunks the content using markdown-aware boundaries (respecting code blocks, headers, paragraphs)
2. Generates vector embeddings via the configured AI provider
3. Extracts and assigns tags using LLM structured outputs (if auto-tagging is enabled)
4. Builds semantic edges to other atoms based on embedding similarity

This pipeline is fire-and-forget from the caller's perspective — the caller receives the saved atom immediately while embedding/tagging runs in the background, with progress reported via callbacks.

Tags form a hierarchical tree. Auto-extracted tags are organized under category parents (Topics, People, Locations, Organizations, Events). Tags serve as both organizational structure and scoping mechanism for wiki generation and chat conversations.

Wiki articles are LLM-synthesized summaries of all atoms under a given tag, with inline citations linking back to source atoms. They support incremental updates — when new atoms are tagged, only the new content is sent to the LLM to integrate into the existing article.