Sovereign AI

Executive summary

Using a cloud AI service typically means transmitting prompts and related context to infrastructure controlled by another organization. That can create privacy, security, continuity, and governance concerns when the information is sensitive. Local systems can reduce some of that exposure, but they introduce their own limits in capability, maintenance, usability, and verification.

UNA explores this tradeoff. Its prototype architecture distributes work across two Apple Silicon machines on a local network and separates reasoning, utility services, memory, and user-facing functions. The design goal is to reduce unnecessary data movement while making authority and model roles inspectable. That goal is not yet equivalent to independently verified zero egress, complete mediation, or production readiness.

This paper makes three narrower arguments: dependence on remote AI infrastructure can create strategic risk; consumer hardware can support meaningful local AI workloads; and personal AI needs enforceable boundaries, evidence, and human authority rather than privacy promises alone.

The important question is not only what an AI can do. It is who controls the system, the data, the authority, and the conditions under which it acts.

The challenge: you don't own your AI's thoughts

When a founder drafts a pitch with a cloud assistant, that strategy touches the vendor's servers. When a therapist structures session notes, those notes traverse third-party infrastructure. In each case the user trades sovereignty for capability.

Encryption, retention controls, contractual promises, and training policies all matter, but none answers the whole governance question. A user also needs to know which systems process the data, which actions are authorized, what evidence is retained, and how failures are contained. Local execution can reduce exposure, but it does not automatically guarantee privacy or trustworthy behavior.

The architecture: distributed cognition, governed

Two nodes, one brain

UNA runs on two machines. The edge node handles API routing, file monitoring, voice synthesis, and user-facing services. The primary brain runs all reasoning, knowledge synthesis, and analysis. Routine operations stay local on the edge node; compute-intensive reasoning routes to the brain's large unified memory: the same specialization principle edge-computing research relies on, applied to cognition.

The governed boundary: why it matters

Lightweight local models can support text-to-speech, embeddings, and utility functions. The architecture assigns those models bounded roles and is intended to exclude them from the primary reasoning path. Configuration declarations and tests provide evidence about that separation, but they do not by themselves prove complete mediation across every loader, tool, memory path, and runtime entry point.

The reason is practical: every model or service that can alter a reasoning pipeline can affect the final result. A trustworthy multi-model system therefore needs explicit roles, constrained interfaces, provenance, and tests that attempt to bypass those boundaries. The present architecture is a candidate implementation of that idea, not final proof that every path is contained.

Three-tier memory

The architecture maps physical storage to different memory functions: unified RAM for active context, NVMe storage for working records and episodic material, and larger-capacity storage for archives and snapshots. This is an implemented design direction, but memory integrity, provenance, retrieval quality, and safe restoration remain separate evaluation problems.

Dual-mode execution

Not every task is the same. Status checks need deterministic output; creative work needs open-ended exploration. UNA runs a dual-mode framework: structured mode enforces task-specific schemas: morphology over creativity at the execution boundary: while open mode permits free-form reasoning behind a relevance gate that catches and retries off-topic responses.

What this means for you

Solo practitioners and founders: your intellectual output is your primary asset. UNA lets you build an AI partner that knows your work intimately across accumulated context, without exposing that knowledge to any third party.

Regulated work: local infrastructure may support tighter data control, but deployment in healthcare, legal, financial, or other regulated settings would still require domain-specific security, privacy, compliance, validation, and operational review. Architecture alone does not establish compliance.

AI researchers: the architecture offers a testable question: can a multi-model system keep reasoning, utility services, memory, and action authority separated under adversarial conditions? The governed-boundary model is a research candidate for that problem.

Conclusion

Cloud-only AI is not the only possible architecture. The UNA prototype shows how consumer hardware, local services, explicit model roles, and governed interfaces can be assembled into a more inspectable personal system. It does not yet prove complete sovereignty, zero data egress, production reliability, or general superiority to cloud systems. Its value is the design question it makes concrete: how much intelligence can remain under meaningful human control, and what evidence would justify trusting it?

A personal intelligence system should remain answerable to the person whose life, work, and memory it serves.

Tom Budd is the founder of ResoVerse LLC, a San Diego AI research company focused on sovereign AI, distributed cognition, and human-AI collaboration. UNA is its flagship system.: Get in touch