I am rebuilding Vicky, the AI assistant I use day to day, so that she works with several knowledge bases at once, one for each area of life, without mixing what is in each. The first stage, which defines where each note should live, is finished and tested. The next challenge is the most interesting one in the project: making sure no information is written to the wrong base. This diary follows the build.
What Vicky is
Vicky is not a product. She is a set of AI agents that work on a vault. It is from this collection, with projects, meetings, decisions and song lyrics, that they draw the context to answer and write. The idea is related to the LLM Wiki that Andrej Karpathy described in April: instead of searching the raw documents for every question, the AI maintains a collection of interlinked pages and looks after it, cross-referencing, updating and correcting.
Why several vaults
A single vault serves a single life, and nobody has a single life. A company's material should not share a folder with the family's notes. And each area needs its own instructions, vocabulary and tools: an assistant that serves work and home with the same rules serves both badly.
The federated Vicky
That is why the new version, rewritten from scratch (in TypeScript, for those who care), is federated. Each vault is independent, with its own owner and rules, and the same assistant consults several: a personal one, a company one, one for each project. In practice, it is a federated LLM Wiki.
The case that excites me most is that of someone starting a new job, connecting the company's shared vault and having, from the first day, an assistant that has already read all the documentation, outdated parts included, and that helps the person get their bearings.
What is already in place
Part of the protection already exists, and it is deterministic: it works without depending on anyone, person or AI, getting it right. A manifest, a configuration file, says which device holds which vault. And the restrictions work in layers: a rule set at the top level can be tightened in a vault, never loosened.
The challenge: the boundary
Each new note has to be written to the right vault, and the AI itself does this triage. By design, the triage does not work alone: behind every decision by the model there will be a safeguard that works even when the model is wrong. Designing these safeguards is the work of the coming weeks.
How we are testing
Everything runs first in test vaults, filled with people, companies and meetings generated for the purpose, with no real data. The agents use these vaults as they would use the real ones, and every error that appears becomes a new rule. The next entries will describe what the tests taught us.
I am publishing this diary openly because Vicky is no secret, and because in this field the ground shifts every week. Ideas for this boundary are welcome: send them through Instagram, @angrio.us.