Shared facts & privacy

A family agent holds the household's life in one shared set of facts that everyone contributes to and everyone can read. This is the heart of the family layer, and it's where the privacy model gets real. The short version: a family is a transparent unit, and privacy is a separate agent, not a hidden room inside the family one.

One set of facts the whole family writes

A solo agent keeps a flat set of durable facts — the things worth remembering about you and your life. A family agent keeps one shared set of facts for the whole household — and every member's private conversation reads from and writes to it. Something the agent learns about the family's week in one person's chat is there when it talks to anyone else.

The one per-member exception is what the agent knows about you as an individual — like what to call you. Everyone has their own. So the model is simple:

• Who someone is → their own personal notes.
• What's happening with the family → the one shared set of facts.

Earlier versions of the family layer had a separate, gated "should this go in the family memory?" step and a second family file. We retired both. One shared set of facts everyone can open is simpler to reason about, cheaper to run, and truer to what a family agent is — an agent that knows your household, not one keeping ledgers on each of you.

A transparent unit

A family agent treats the household as a unit that can see itself. A parent can read what the family agent knows; the shared knowledge — the family facts, the calendar, the notes — is visible to everyone in the household through the Context tab. That's the deliberate stance: families work on trust and visibility, not walls between the people in them.

What that means concretely depends on how someone joined, because the join method is the natural line:

• Code + PIN members (typically a kid on a shared device) are visible to the household's adults. An admin can look at what the agent knows about them and their conversation — the access a responsible parent should have over the AI a young kid is talking to, satisfied by the architecture rather than by a "trust us" promise.
• Invited adults (joined by their own email account) get a private layer. Another adult — even the owner — can't read their personal notes or their private conversation. Peers don't get to read each other.

Privacy is a separate agent, not a wall

If someone in the household wants space that's genuinely their own — thoughts they don't want shared into the family facts — the answer isn't a hidden compartment inside the family agent. It's a separate solo agent, which is private by the same rules every solo agent is. Keeping the family agent a transparent unit and pointing private needs at a private agent is cleaner, and easier to reason about, than a family agent with secret rooms.

The safety floor still comes first

There is one carve-out, and it's a safety one. If a child discloses something that triggers the constitution's safety floor — a crisis, harm, the serious stuff — that's handled by the safety layer, which takes ultimate precedence. It is not quietly written into the shared family facts, because the architecture can't assume every adult in a household is a safe audience for it. The transparent-unit default bends here, on purpose, in the one place it has to.