I don’t have much local model experience, but the main bottleneck was the agent’s loop and good tool calling to browse the notes. The topic parsing doesn’t require that much intelligence, and the linking of topics was done with co-occurrence and graph algorithms, outside the LLM.
@hmans.dev Thanks for github.com/hmans/beans! It pairs beautifully with github.com/raine/workmux, and together they've completely changed my workflow
I've written a blog about how it works and what I learned pieterma.es/syntopic-rea...
I let #Claude loose in a library of 100 books to see if it could find interesting connections.
One of my favourites is trails.pieterma.es/trail/invisi..., about the gradual crumbling of structures, systems, and institutions.
