Voical Whitepaper · Lucid Systems LLC · May 2026

Sovereign AI
on Your Desk

How we built a complete AI operating system that runs entirely on your hardware — no cloud, no subscriptions, and no data ever leaving your device.


01 — The Case for Owning Your Intelligence

Most AI today is rented.

You pay a monthly fee for access to someone else's models running in someone else's data centers. Your prompts, your voice, your documents, and your private thoughts are sent across the internet, stored, and used to train future versions of those models. If the company changes its terms, raises prices, or simply decides you are no longer a welcome customer, your access can be revoked on a Tuesday morning.

This is not intelligence you own. It is intelligence you lease.

Voical exists to change that.

We built a complete local AI operating system that runs entirely on your own Apple Silicon hardware. No cloud. No subscription. No data ever leaves your trusted devices unless you explicitly choose to open the bridge.


02 — What Voical Actually Is

A beautiful, fast, native
Mac application.

Thalamus Mesh — M4 Conductor with M2 Mini, M3 Trusted Peer, Edge Node

Thalamus Mesh turns your trusted hardware into a single distributed intelligence.

Everything is paid for once. Everything runs on your hardware. Everything belongs to you.


03 — The Architecture

Three progressive transport
layers. Each is optional.

Layer 1 — Git (always active, fully air-gapped)

A git repository acts as the durable, conflict-tolerant message bus. All tasks, results, verdicts, chat messages, and memory anchors are stored as versioned Markdown files.

Layer 2 — LAN (port 47251 + mDNS)

When devices are on the same Wi-Fi, real-time notifications dramatically reduce latency while content still flows through git.

Layer 3 — Internet Bridge (opt-in, content-free relay)

When explicitly enabled, a lightweight WebSocket relay carries only task IDs and notifications — never content, never model output, never private data.

Three transport layers — Git, LAN, Internet Bridge

Each layer adds speed without changing protocol semantics. Remove the internet layer and the mesh continues to function perfectly.

This design means Voical works perfectly offline, becomes dramatically faster on LAN, and can optionally reach trusted peers across the internet without compromising the privacy model.


04 — Memory & Long-Term Intelligence

The single biggest limitation
of local LLMs is context.

Voical solves this with a three-layer memory system:

Voical three-layer memory system

Voical's three-layer memory system gives the agent coherent long-term intelligence without ever exceeding the context window.

Live Context — The active conversation window (gated by available RAM: 8192 tokens on M4-class machines).

Session Compaction — When the context approaches capacity, Voical automatically (or manually) compresses the conversation into a dense, natural-language summary using a general-purpose model. The summary is treated as ground truth in the next session and is persisted to disk so it survives restarts.

Memory Anchors + RAG — Every compressed session and manually saved note becomes a versioned "anchor" in ~/voical-memory/. A live semantic index (plus a historical "Pulvinar" index that reaches back through git) lets the agent retrieve relevant context from weeks or months earlier.

Smart chunking, cosine relevance thresholds, incremental indexing, and an anchor eviction policy keep the system fast and sustainable even after hundreds of sessions.

The result is an AI that feels like it actually knows you — not just within one chat, but across time.


05 — Resilience & Real-World Use

Designed for the real world.

Voice cloning, the proactive agent, document writing, and code assistance all work offline. The mesh and memory systems continue to evolve locally.


06 — Trust, Security & Identity

Trust rooted in three things.

01

Hardware boundary

Your data never leaves your trusted devices by default.

02

Cryptographic identity

Every node has an Ed25519 keypair. All messages and results are double-signed.

03

Explicit consent

Dangerous tools and internet actions require user approval with configurable levels — ask always, session, or safe-only.

Voical trust boundary — concentric rings showing data containment

Voical's strict trust boundary. Data stays inside your trusted hardware by default.

There is no central server, no telemetry, and no ability for us to revoke access.


07 — Honest Limits & Roadmap

We're not trying to beat
frontier cloud models today.

Those models are still stronger on certain creative and reasoning tasks.

What we are better at is ownership, privacy, speed on your hardware, long-term memory, and multi-device collaboration.

Current roadmap priorities include:


08 — Why This Matters

We are at an
inflection point.

AI is becoming one of the most powerful tools humanity has ever built. The question is no longer whether it will be transformative — the question is who will control it.

Will your intelligence be rented from corporations that can change the rules at any time?

Or will it belong to you — running on hardware you already own, under rules you set, with data that never leaves your desk?

Voical exists so the answer can be the second one.

You don't have to rent your mind.
Welcome to owned intelligence.