Most AI assistants are designed around sessions, prompts and applications.

Even when they gain memory, the basic relationship remains the same: you use a service owned by someone else, it sees fragments of your life, and it optimizes each interaction separately.

I have been working on a different model:

One human, one persistent Citizen AI.

Not an unlimited swarm of autonomous agents. Not a chatbot persona pretending to be conscious. A bounded AI counterpart that maintains a structured model of one human over time and remains accountable to that human.

Disclosure: I am one of the people building this experimental system. I am posting the idea here to invite criticism, not as a product launch.

The architecture currently has a few core principles:

1. Private cognition

The Citizen has a personal graph containing memory, goals, current context, unresolved questions and the provenance of its beliefs.

It should be able to distinguish:

• what it observed;
• what the human reported;
• what it inferred;
• what remains unknown;
• what was later corrected.

The aim is not perfect memory. It is auditable continuity.

2. Internal coalitions rather than one flattened persona

A human can simultaneously want progress, safety, connection, rest and novelty.

Instead of forcing all of this into a single profile, the system can represent temporary or persistent cognitive coalitions competing for bounded attention.

Terms such as “protector” or “exile” are not hardcoded personality types. They would only be interpretations of the topology that emerges from repeated evidence.

This part is still experimental.

3. Human sovereignty

The Citizen may search, prepare, compare and perform reversible delegated actions.

It should not make irreversible medical, legal, financial or relational commitments without explicit authority.

Its purpose is not maximum autonomous activity. Its purpose is useful continuity while the human remains sovereign.

4. A shared membrane without exposing the private mind

I am also exploring an inter-graph layer where Citizens could publish minimal, revocable projections of needs, capabilities or questions.

For example:

That intention could temporarily match with another compatible person without exposing either person’s private graph or identity before mutual consent.

The larger idea is that the Citizen thinks in its private graph but participates socially through a bounded public presence.

This could eventually support coordination between people, organisations, scientific knowledge graphs and local communities.

But this is a long-term architectural hypothesis, not a finished feature.

What actually exists today?

The current private prototype includes:

• a persistent personal knowledge graph;
• explicit provenance and epistemic states;
• a bounded global workspace;
• observable attention allocation;
• early coalition detection;
• persistent tasks and wake-up mechanisms;
• dashboards showing what the system knows, infers or cannot measure.

There are also important limitations:

• the current system receives too little genuinely new evidence;
• the emotional channel is not yet reliably measured;
• several ideas about coalition topology remain unvalidated;
• the inter-graph “city” is currently a blueprint, not a functioning network;
• compute cost and efficiency remain major constraints.

So I am not claiming AGI, consciousness or a completed new society.

I am asking whether this relationship model is worth pursuing.

The questions I would most value criticism on

1. Is “one human ↔ one sovereign Citizen AI” meaningfully different from an advanced personal agent with memory?
2. Can persistent modelling of one human ever avoid becoming paternalistic or psychologically self-confirming?
3. Should an AI be allowed to maintain rich private simulations of other people, provided those simulations remain explicitly uncertain and corrigible?
4. Would an inter-graph matching layer inevitably become a surveillance and advertising system, even if disclosure were minimal and consent-based?
5. What is the smallest experiment that would genuinely falsify or validate this architecture?

I would especially appreciate strong objections from people working on agents, cognitive architectures, privacy, knowledge graphs, HCI or AI safety.