Blog/2026-07-01/The Everything Store
Having a single store so I can query my second mind of things I've written down has always been something I've wanted. That's evolved in various ways. At one point, I had these as just plain text files on my computer, then I had some promoted to my wiki, and so on. But when I added semantic search it struck me that the same machinery works for many other things.
So I built for myself the simplest possible incarnation of this:
- A SQLite vector database
- A simple HTTP server in front of it
- Reused the Wiki semantic search's embedding model (bge-en-base)
And the results are honestly quite impressive. With a couple of skills for Claude Code, Codex, and pi that allow access to the task list and the KB, all modern models search on it quite effectively and inform themselves of what to do and how to work. The data model is pretty trivial:
- Containers are the bucket to put things in: the agent, or home things, or travel things, or one of my projects
- Documents are a tree-like structure
My friend Armaan pointed out that this is pretty much the bucket and key structure that S3 uses and so I must have unconsciously chosen a similar format. Modern agents are pretty good at managing this kind of thing so I try to stay as unopinionated as possible except for the container/bucket name after which they organize things how they see fit.
I have a UI where I can browse through if I'm interested but for the most part, it's mostly agent managed and I talk to it with either the Telegram bot or through one of the coding harnesses. The 512-token limit is quite annoying and I simply didn't deal with it at all, though Armaan in his implementation did the naïve thing that works: a key under the path can have multiple segments each of which is appropriately embedded. Pretty neat. I haven't adopted it yet but I surely should.
Now, the thing I'm wondering about is whether I should simply make this the real Everything Store and put everything in there and have other programs be views into it, but that seems like it will run into common data-modeling issues and so on. E.g. I don't have any images or anything in here, so the agents store them on Google Drive and then reference them here for later. I think in the long-term the impedance mismatch between the manner of use from various programs will dominate so I have our EMR system (which carries X-rays, events, immunization) store its own data and just use adapters/skills to access that, but information on the EMR system is held in the Everything Store.
Honestly, it's surprising and lovely how few things one has to plug into each other to get fantastic results here. Agents naturally use the information in the KB without much work required. In the past, something like this would be pure science-fiction but now an agent can look up our appointments, travel plans, flights, and put it all together to be an actually helpful assistant.
