What the Mind Mesh Actually Does

In my last post I wrote about moving my ingestion from Notion to Obsidian, with Claude Code on top to process the raw input. I ended on the idea of an intelligent layer sitting on top of those files, without really knowing what it would look like. It was a direction more than a plan. Over the past few months it has become a lot clearer, and I now have a first version of it running.

What the layer actually is

Everything I read, watch, or listen to lands in the vault as a plain file. But a file is only a record of what one source said, and a pile of them is not a view on anything. The layer turns them into a map.

The agents turn the subjects that keep coming up into nodes, and the relationships between them into links. Each node holds a set of claims about that subject, and as new material arrives they file each source under the claims it speaks to. The map is never rebuilt from scratch. It updates incrementally, one source at a time, with each new piece of material folded into what's already there.

That's what makes the map worth reading. Convergence and divergence become visible on their own. When several sources independently point the same way, it shows, and so does one source disagreeing with the rest. Neither is visible when the sources sit in separate files read months apart.

What I still do

My reading is still the bottleneck, but what I'm reading has changed. The agents do the first pass through the raw material, and I read what they write and propose. I still go back to the sources when something needs a closer look, and I review and approve everything before it gets written down, which is how I make sure I understand what the material says rather than only holding a map of it.

The agents do more than get through the material, though. They surface things I wouldn't have known to look for, either because they sit outside what I know or because I read straight past them. And their ability to reference what's already in the vault and connect one thing to another is well beyond what I can do on my own, since searching my own memory is neither exhaustive nor particularly reliable.

I'm not sure, but I suspect AI eventually gets better at this than I am. If that happens, the whole middle layer, everything between raw input and whatever output I need, can be offloaded, and my staying in the loop would be not just a bottleneck but a destructive one. Perhaps that's what AGI looks like from this angle.

Where this goes next

There are at least two things I want to do next. The first is iteration, since the workflow still has rough edges and most of the near-term work is smoothing them out.

The second is an app. The map lives in files, so today I read them one at a time. Obsidian can show how they link together, but that isn't really what I need, and as the map grows I want something shaped around it.