Claude Code + Obsidian: My family’s AI brain

Scott Farrar says he didn’t want a “complex problem” to live only inside a chat thread where the model’s memory and drift feel opaque, so he built a local system in Claude Code that reads/writes Markdown files as a transparent record of what the AI “knows.” That lets him start new sessions while retaining continuity through the same files.

In his car insurance example, Scott describes how the system summarizes documents he drops into a folder, tracks quotes, identifies blockers, and generates a concrete next step (like emailing an agent with specific questions). He contrasts this with previously postponing the task or relying more heavily on an agent without reading details.

Scott keeps structured household context in an Obsidian vault—appliance serial numbers, warranty info, device model/OS notes, photos of labels—and then queries it conversationally (e.g., “What’s going on with the fridge?”). He says Claude can look up the relevant files and respond using that stored context.

Claire raises a “type A vs. type B” organization spectrum; Scott responds that he can dump unstructured info and rely on the coding agent to structure and maintain it, including enforcing organization rules and using a linting script to keep outputs from becoming too verbose. The system blends flexibility (dumping data) with guardrails (structure enforcement).

To collaborate with his wife, Scott tried a shared Dropbox folder for the vault and considered a GitHub repo approach; he notes Dropbox can create sync conflicts if both edit simultaneously, while GitHub brings overhead. He ends up using a hybrid: Dropbox for most content and GitHub for certain scripts.