MERCURY NOTES

Almost certainly the most technically interesting part. When it looks at your note, it gets an embedding of that note and finds all the other notes with similar semantic content and looks at the tags on those notes. Then it uses a few other heuristics like how long it's been since those notes have been edited and comes up with its best guess of what tag you want based on tags of similar notes. When you first start using the tool, it doesn't really know what subjects you gravitate around. But as you use it more, it learns. What's really important is that the tag's name doesn't have to have anything to do with the content because it's using other semantically similar content to figure out tags rather than the tag name itself.

Notes get tagged (e.g., 'D&D' + 'brainstorming'). For each tag+category combination, the LLM reads all notes with those properties and regenerates a summary document containing all component notes. The summary is beautifully formatted and nicely synthesized, with original notes linked as references. Summaries are cached—when you come back, it's the exact same summary. The summary page shows it's 'dirty' when new notes are added, and you can manually regenerate. In production, could have a batch job that regenerates any dirty summaries older than one day.

I was really proud of the way it converted collections into to-do lists. If I mention like, 'Oh, here's a laundry list of things I need to do today,' it will do a detection of what type of note. If it detects that it's a collection or a to-do list, it will turn it into structured data and save it as such instead. Your to-do list suddenly becomes this actual to-do list rather than just a collection of random scribbled notes.