Local Companion AI

01

The challenge

A useful companion needs continuity, perception, and agency, but an always-aware system also creates hard constraints around battery, privacy, storage, and user control.

02

The approach

1. Target a modern Android phone as the eventual home for the model.
2. Separate short-term context, durable memories, and learned preferences.
3. Study self-evolving memory patterns without allowing silent behavioral drift.
4. Use memory loss on power failure as both a technical constraint and design premise.

03

Where it stands

This remains a research and architecture project. Current work is defining a realistic minimum companion loop before choosing the local model and device runtime.

04

Next useful moves

• Define the smallest useful perception-memory-response loop.
• Benchmark candidate small models on available Android hardware.
• Write explicit privacy, deletion, and user-override rules before persistence work.