Hal Universal

A private on device AI assistant with transparent memory and model control. Fast, expressive, and built entirely for your device.

Hal Universal

Private. Powerful. Personal.

Hal Universal is an on-device AI assistant for iOS, iPadOS, and macOS. It combines Apple Foundation Models with the fully local MLX Phi-3 model to deliver fast, expressive, and private conversational intelligence built for Apple platforms.

When Hal runs with Apple Foundation Models (AFM), inference can occur entirely on your device when offline, but may occur in Apples Private Cloud Compute when connected to Wi-Fi or cellular. Private Cloud Compute means your prompts are encrypted in transit and processed only in short-lived, non-persistent memory on Apple-controlled servers that meet the same security guarantees as on-device execution. Apple states that no data is retained after processing.

For users who prefer complete local operation, AFM runs 100 percent on-device when the device radios are disabled.
iPhone / iPad: Turn on Airplane Mode and turn Wi-Fi off
Mac: Turn Wi-Fi off and disconnect Ethernet

When offline, AFM is fully local. When using Phi (MLX), Hal is always fully on-device, regardless of connectivity.

---

Overview

Hal Universal was built for two purposes: to be a thoughtful, private companion for reflection and creativity, and to help people understand how modern LLMs actually work.

Hal exposes the mechanics of memory, recency, decay, retrieval, and tokens so the black box becomes transparent while keeping everything on your device.

---

Key Features

Local-First Design
Phi-3 always runs fully on-device. AFM is fully local when offline.

Timestamp-Aware Memory
Hal uses recency, semantic weighting, and optional half-life decay to determine what matters. All memory lives in a private, on-device SQLite database.

Power User Controls
You can tune how Hal thinks, including memory depth, recency weighting, semantic importance, half-life, retrieval size, similarity settings, and temperature. All setting changes appear directly in the chat for full transparency.

Detailed Token View
Token usage is displayed for every message in the detailed view.

Cross-Platform Parity
Hal now behaves and looks the same across iOS, iPadOS, and macOS, including a resizable Mac window.

Improved Usability
Keyboard dismissal is smoother on iPhone and iPad, and document upload status is clearer.

---

Troubleshooting and Maintenance

Reset (DB Nuke)
This clears Hals local SQLite database and all conversation memory. Models remain installed, and nothing leaves your device.

Model Removal and Redownload
Model caches can be managed in Settings > Manage Models.

---

Privacy Policy

Hal Universal does not collect, store, or share any personal data. All inference, memory, and processing occur locally on your device. The app does not transmit your data externally.