Architectural Decisions & Zero-Cloud Constraints
Every design choice in AXIOM OS prioritizes strict local privacy, minimal battery drain, and deterministic accountability over cloud convenience.
Modern productivity agents rely heavily on cloud APIs to process context. This fundamentally compromises privacy when dealing with screen content, private messages, and personal work habits. AXIOM OS enforces a strict zero-cloud constraint.
No OpenAI, Anthropic, or external credentials required. The system works completely air-gapped.
All intelligence executes directly on local CPU/GPU silicon using quantized GGUF weights.
Telemetry never leaves the 127.0.0.1 interface, enforced with operating-system socket blocks.
Running an LLM constantly in the background can monopolize system resources. We evaluated several open-weights models before settling on the 3-billion parameter class for everyday background execution.
| Candidate Model | VRAM Footprint | Reasoning Speed | Engineering Verdict |
|---|---|---|---|
| Llama-3-8B | ~6.0 GB | Medium | High reasoning quality, but excessive memory drain on ultraportable laptops. |
| Qwen-2.5-3B (Selected) | ~2.5 GB | Sub-second (Very Fast) | Optimal balance: Fits in unified memory with negligible thermal footprint. |
| Phi-3-Mini | ~2.2 GB | Fast | Fast inference, but exhibited context truncation on multi-hour timeline summaries. |
A critical design decision was separating the factual measurement of behavior from the AI interpretation. The Go daemon writes strict facts to SQLite (such as "Chrome was open on youtube.com for 5 minutes"). The LLM is only given read-access to this data to form opinions.
This guarantees the AI cannot hallucinate past user actions; it can only judge recorded, immutable ground truth.