Features

Six pillars, zero servers.

What Silver Prime actually does today — and what's honestly on the way. Every card is tagged shipped, in progress, or planned.

01 · Inference
In progress· 3.1 → 3.3Free + Premium

Hybrid LLM routing

Cloud-only today via your Groq key. Phi-3-mini on-device is the next major milestone. When both exist, a small classifier will pick where your message goes based on length, keywords, and your preference.

  • Groq llama-3.1-8b-instant via your key — shipped
  • Anthropic / OpenAI / Gemini adapters — planned after Stage 3.3
  • Phi-3-mini-4k-instruct Q4_K_M, ~2.2 GB first-run download — planned Stage 3.3
  • Heuristic complexity score: length, keywords, context turns, images — planned
  • Override in settings: Always local / Always cloud / Smart (default)
02 · Privacy
ShippedFree + Premium

Privacy by architecture, not policy

There's no backend to leak because there isn't one. API keys sit in the Android Keystore. Conversations sit in SQLCipher. Biometric data, if you ever opt in, never leaves the device. Your conversations never leave the device — local inference, local keys, local memory.

  • No server-side auth, no analytics, no telemetry — shipped
  • API keys wrapped by Android Keystore (hardware-backed) — shipped
  • Conversations in SQLCipher-encrypted SQLite — shipped
  • Face embeddings never transmitted (if you enrol later) — planned Stage 4
  • One-tap wipe for face + voice data — planned Stage 4
03 · Persona
Planned· 3.5Free + Premium

Renameable assistant

Default persona will be "Prime." Change the name in onboarding or settings and every system prompt, every notification, every greeting follows.

  • Default: "Prime" (mirrors Siri / Alexa defaults)
  • Custom name persists across sessions + backups
  • Persona string injected into system prompt
  • Voice wakeword will adapt to the chosen name (after Stage 3.4 voice lands)
04 · Streaming
ShippedFree + Premium

Streaming chat

Token-by-token replies via SSE. Typing cursor pulses while the model thinks. Abort mid-generation when you already got what you needed.

  • XHR-based SSE parser (Hermes-safe on RN 0.74)
  • AbortController cancels in-flight streams
  • Assistant turn persisted as it streams, not after
  • "+ New" session button resets + aborts in one tap
05 · Memory
ShippedFree + Premium

Conversation persistence

Every session, every turn, stored locally. Launch the app six weeks later, your last chat is still there.

  • WatermelonDB + SQLCipher schema
  • conversation_sessions + conversation_turns tables
  • 50 most-recent turns hydrated on launch
  • Provisional title from the first user message
  • Archive, restore, export — all local
06 · Self-improvement
Planned· 4 + 6 + 7Premium

Recursive self-improvement

Ask the assistant to do something it can't. It will draft a capability, run it against fixtures in a sandbox you can inspect, and — once you approve — the capability becomes a permanent part of your assistant. JS-layer self-update is free and local: the assistant writes a new JS module, the Hermes isolate runs it, and if you confirm it works, it's written to the app's private storage and loaded on the next launch. No network required, no APK reinstall. Native features — new camera hooks, accessibility permissions, system integrations — live in the APK's native binaries, which Android signs and freezes. For those, the Pro channel (Stage 6) rebuilds a signed APK and prompts you to install the update. You always confirm. A skill marketplace (Stage 7) lets you publish capabilities you built — other users buy them for $SOL on Solana with marketplace-enforced royalties back to you.

  • JS module generated via your API key (Groq / Anthropic / OpenAI / Gemini)
  • Static-analysis gate: no eval, no require, no __proto__, no host-app access
  • Runs in Hermes isolated runtime with fixtures before promotion
  • User approves every capability; 30-day one-tap rollback
  • Optional Pro channel (Stage 6) rebuilds signed APK for native features
  • Optional marketplace (Stage 7) lists skills on Solana for $SOL

Privacy FAQ

Where does the on-device model live?

The default install places the ~2.2 GB Phi-3-mini weight file in Silver Prime's private app storage, excluded from Android Auto Backup. Developer-mode users can point the app at a sideloaded file in shared storage — in that case the model file itself isn't private, but since Phi-3 is a public open-source model from Microsoft, its location reveals nothing about you. Your prompts, conversations, keys, and memory remain private on-device in all modes.