Local-First System Specification

System Topology & Daemon Architecture

AXIOM OS operates entirely on the local machine. There is no cloud telemetry, no remote inference, and no external API dependencies. The architecture is strictly separated into three layers: Collection (Specter), Storage (SQLite), and Intelligence (Ollama + Qwen).

1. Top-Level Local Topology
Air-Gapped Flow

Strict process boundaries guarantee that telemetry stays local. The Go daemon writes to SQLite; the local Ollama LLM queries SQLite on demand to evaluate context.

2. Specter Daemon (Sensory Organ)
Go Core

Written in Go for minimal overhead, the Specter daemon is the sensory organ of AXIOM. It continuously monitors window focus, active processes, and system state. It does not analyze text; it merely records deterministic facts.

Process Name

specter.exe

Memory Footprint

< 15MB RAM

Sampling Interval

1Hz (1 poll / sec)

3. Local Memory Layer (SQLite Schema)
Immutable Ledger

All telemetry is funneled into a local SQLite database. This acts as the single source of truth. The AI layer cannot modify the raw telemetry, ensuring an immutable ledger of behavior that prevents the LLM from hallucinating past user actions.

// TELEMETRY SCHEMA
CREATE TABLE telemetry (
    id INTEGER PRIMARY KEY AUTOINCREMENT,
    timestamp DATETIME DEFAULT CURRENT_TIMESTAMP,
    active_window TEXT NOT NULL,
    process_name TEXT NOT NULL,
    category TEXT DEFAULT 'UNKNOWN'
);
4. Local AI Interpretation (Ollama + Qwen)
On-Demand Infer

The intelligence layer runs on demand, querying the SQLite database for recent behavior contexts. Using small, efficient local models via Ollama (such as Qwen 2.5 3B), it cross-references recorded actions against predefined user policies to generate Focus Scores and deliver gentle or blunt contextual coaching.

Evidence: Local-First Privacy & Offline AI Execution
0 BYTES EGRESS
$ specter start --daemon
[+] Telemetry ingest started locally. Network: DISABLED (0 bytes egress)
[+] SQLite WAL store: opened at ~/.axiomos/telemetry.db
[+] Ollama inference engine loaded in memory (qwen2.5:3b)
[✓] Policy Engine: Running locally without cloud APIs
[✓] Loop interval: 1Hz, memory usage: 14.2MB