Guglielmo Brain

The problem: an assistant is only as good as what it is allowed to remember

An AI assistant that forgets between sessions is a stranger every morning. One that remembers everything is a liability, because the same store now holds a project decision and a medical fact. A useful personal system has to remember continuously and forget selectively, and the selection is the hard part.

The easy version of this is a document the assistant reads. It fails immediately, because the moment the assistant can also write, every piece of context it captures has to be classified before it lands somewhere, and classifying it wrong is how private information ends up in a professional output.

The system is built around that classification rather than around the storage. What to keep is easy. What to route away from an LLM, and what to never send at all, is the design.

Four privacy levels, enforced before anything is written

Every fact carries one of four levels: public, safe for external professional output; private, internal only; sensitive, requiring explicit review before it can be shared; and secret, which is never sent to an LLM, an index or any external system. Credentials and environment files are always secret, by rule and not by judgement.

This is exactly why nothing on this website exposes a codebase valuation, a location or a personal number: those facts exist in the record, and they are classified below the line that reaches public output. The website is downstream of a boundary that was drawn before it existed.

The four levels are a boundary that a machine can apply consistently, which matters because consistency is precisely what a human classifier lacks at the end of a long day.

The decision that makes it trustworthy: write decisions are deterministic

Before any input is persisted, a deterministic policy decides where it goes: ordinary app state, a durable Brain claim, or a review queue. The decision is not left to the model. A fact that conflicts with an existing claim, or that touches health or finance, is routed to review rather than written automatically.

Letting the assistant decide, per input, whether something is safe to store is the failure mode, because it is unpredictable in exactly the cases that matter. A fixed policy that runs before persistence makes the behaviour reviewable: you can read the rule and know what will happen, instead of hoping the model chose well.

The clearest example is the medication guard: if the assistant is asked to create a reminder to take a medication, and the record already contains a claim that the medication is not being taken, the write is forced into review rather than executed. A contradiction between a new instruction and an established fact is treated as a signal to stop, not a value to overwrite.

Why it exists at all

Every page on this site is checked against this record before it is published, which is what keeps the public claims accurate and consistent. The system's job is to be the single source of truth that a professional output can be derived from without inventing anything.

A portfolio drifts from reality the moment it is written by hand, because the hand does not remember what changed. Deriving public output from a maintained record instead means the claims stay true as the work moves, and the boundary between what is shown and what is kept is drawn once, in the system, rather than re-decided every time something is written.

Questions

What makes this different from giving an assistant a long memory?

The hard part is not remembering, it is deciding what may never be stored. Every fact is classified into one of four privacy levels before it is written, and secret data such as credentials is never sent to an LLM at all. A long memory without that boundary is a liability, not a feature.

How does it decide what to store?

A deterministic policy runs before anything is persisted and routes each input to app state, a durable claim, or a review queue. The model does not make the call. Anything touching health or finance, or anything that conflicts with an existing claim, is sent to review rather than written automatically.

What is the medication guard?

A concrete example of conflict handling: if the assistant is asked to remind the user to take a medication, but the record already states that medication is not being taken, the write is forced into review instead of executed. A contradiction between a new instruction and an established fact stops the write rather than overwriting the fact.

Facts on this page verified 2026-08-08