Brain

A personal knowledge system where spoken brain-dumps and dropped files become Markdown notes that stay the source of truth, readable in Obsidian and queryable in natural language. Local embeddings index the notes. It answers chronology-aware questions like “what occupied me last week”, extracts the tasks hiding inside a ramble, and asks before it acts on any of them.

The problem Link to heading

Note apps store text but cannot answer questions about it, and cloud assistants that can answer want the whole private archive uploaded first. Neither reads a voice transcript and turns it into what actually needs doing.

What I built Link to heading

Built a Python system where capture, extraction, retrieval and dispatch are separate stages over plain Markdown. Embeddings run locally against a LanceDB index; the chat layer routes through a fallback chain of cloud backends so a low-RAM machine is never forced to load a local model. Extracted actions are proposed, not executed, and reach Apple Reminders or Telegram only after approval.

Implementation Link to heading

• Your files stay yours: Markdown notes in a normal folder are the source of truth. The index is derived and can be rebuilt from scratch.
• Retrieval that knows about time: Local embeddings plus contextual retrieval and generated maps of content, so questions about a period answer from what was actually written then.
• Voice to task, with a human gate: Transcripts are cleaned, actions are extracted with due dates, and each one is approved through a Telegram bot before it becomes a real reminder.
• Claims get checked: Factual claims pulled from entries are verified against live web sources and recorded with a verdict, instead of being trusted because they were said out loud.
• Built to be maintained: 111 test modules, an entity and person graph, and 15 shipped phases planned as a decision register rather than a backlog.