Skip to main content

Why The Device Agent

Most personalization today is a cloud profile, a fine-tuned model release, or a prompt full of user text. Edge takes a different contract: the device is the Agent, the app is the carrier, and learning is local runtime state that restores only when compatibility checks pass. This page explains why — and what the quickstart actually proved.

The Product Shift​

A model release cannot follow each user. Cloud profiles replay private text into every request. What a personal assistant actually needs is: the base model package stays stable, and each user's device keeps learning the user — locally, removably, auditable.

Edge calls the local learning artifact a Neural Imprint. It is restored only when compatibility checks pass, it is removable local data, and it does not require putting private profile text into every prompt.

What Actually Happens In A Learning Run​

When you ran the quickstart, this is what happened:

  • The Agent learned from a local synthetic signal you inspected first.
  • RPP self-learning produced a local learning representation.
  • Neural Imprint restored that state into a compatible model session.
  • Restore passed through model, tokenizer, runtime, and tool-schema compatibility checks.
  • The tool policy showed which local tools are appropriate and which actions are out of bounds.
  • If compatibility fails at any point, the product keeps the base-model path.

This is not LoRA, SFT, prompt stuffing, or cloud personalization.

Is This LoRA Or SFT?​

No. LoRA and SFT are useful when you intentionally want a trained model or adapter release. That requires compute, data curation, release packaging, rollback, and regression evaluation. Neural Imprint is a different contract for per-user device learning: the base model package stays stable, and local learning state is restored only when compatibility checks pass.

Is This Prompt Stuffing?​

No. Prompt stuffing repeats profile text or instructions inside every request. That consumes context budget and replays private state. The after-learning chat restores local runtime state from a Neural Imprint receipt, then handles the current message through the normal generation path.

What Did The Agent Learn In The Quickstart?​

Only the synthetic signal you inspected: risk boundary, cash-flow context, trust boundary, and expected local tool policy. In a real product, those signals would come from app-approved local settings, explicit user preferences, and user-visible corrections. The carrier app owns that policy.

Does The Exported App Contain The Mac Learning Result?​

Not by default. The export does not automatically package the Mac learn_receipt.json or its Neural Imprint artifact into the app. That is deliberate: a user's learned state should be owned by the device/carrier lifecycle, not silently baked into a template ZIP. Use the Mac demo to inspect and prove the mechanism; use the exported app to wire the same lifecycle on a real device.

How Does The Phone Learn A New Preference?​

The carrier records a user-approved signal locally, such as a setting, correction, or classified fact. When the app decides the signal is eligible, it starts a device-side Edge Halo job. That job uses the local model session, local tool schema, and bundled RPP A-library to build a new Neural Imprint capsule. Restore is compatibility-checked and fail-closed. The phone does not need to return to the Mac or re-export the app for every new preference.

Does answers_differ: True Prove Production Readiness?​

No. It proves that the restored Neural Imprint artifact is active for a controlled synthetic example and that the answer moved after restore. Production readiness still needs task-specific evaluation, UI controls, deletion UX, and real-device validation.

Where To Go Deeper​