Context Engine

Agent Skill: real-device-qa

The repository also ships a portable real-device-qa skill for implementation-grounded physical-device testing:

npx skills add meharajM/context.md@real-device-qa -g -y

The canonical Agent Skills path is skills/real-device-qa/SKILL.md. It requires a connected iOS device and WebDriverAgent; the skill does not replace those runtime prerequisites.

Capture thoughts locally. Organize them intelligently. Own your memory.

Context Engine is an iOS-first, local-first AI memory system built with React Native and LiteRT-LM.

The app captures typed and spoken thoughts, synthesizes them fully offline using on-device language models, and stores them in a structured context.md memory file.

Why Context Engine?

Most AI memory tools are cloud-first. Context Engine is designed around:

• privacy-first local inference
• offline thought capture
• local markdown persistence
• on-device LLM synthesis
• resilient fallback behavior
• transparent architecture

Even when AI synthesis fails, every non-empty thought is still persisted safely.

Current MVP Features

• Manual typed thought capture
• Push-to-record voice capture
• Local Whisper transcription
• LiteRT-LM on-device synthesis
• Markdown-based memory persistence
• Queue retry + failure fallback system
• Runtime model download + management
• Foreground-only wake-word architecture
• iOS simulator validated MVP

Architecture

User Input
   ↓
Capture Layer
   ↓
Audio / Manual Queue
   ↓
Synthesis Engine (LiteRT-LM)
   ↓
Context Routing
   ↓
context.md Persistence

Core modules:

• src/modules/ContextManager → markdown parsing + persistence
• src/modules/AudioEngine → Whisper-backed recording
• src/modules/SynthesisEngine → LiteRT runtime + queue
• src/core/store.ts → lifecycle + runtime coordination

Tech Stack

• React Native 0.84
• TypeScript
• Zustand
• whisper.rn
• LiteRT-LM
• React Native FS
• Jest + Detox

Verified MVP Status

Validated flows include:

• app launch
• manual capture
• queue processing
• LiteRT synthesis
• model downloads
• persistence fallback
• recording start/stop reliability
• simulator smoke tests

The current MVP is validated primarily on iPhone 16 iOS simulator workflows.

Known Limitations

Current intentional constraints:

• iOS-first architecture
• Android LiteRT/NPU deferred
• No background wake-word
• No lock-screen capture
• Runtime-only settings persistence
• No cloud sync yet
• No hardware trigger integration yet

Local Setup

Requirements

• Node.js >= 22
• Xcode
• CocoaPods
• iOS Simulator

Install

cd ContextEngine
npm install
cd ios
bundle exec pod install
cd ..

Run

npm run ios

Validation

npm run typecheck -- --pretty false
npm run lint
npm test -- --runInBand

Recommended LiteRT Model

Recommended model:

• Gemma3-1B-IT

Expected local model path:

/Documents/models/gemma3-1b-it-int4.litertlm

The settings screen can download and activate the model directly on-device.

Roadmap

Completed

• Offline capture pipeline
• Queue reliability system
• LiteRT integration
• Local synthesis
• Model download flow
• Foreground wake-word architecture
• Release validation

Planned

• Persistent settings
• Real-device performance benchmarks
• Android support
• Encrypted local memory
• Agent synchronization APIs
• Cross-device sync
• Hardware trigger support

Open Source Contribution

Contributions are welcome.

Please read:

• ContextEngine/AGENTS.md
• ContextEngine/implementation/README.md
• ContextEngine/implementation/status.json

before making architectural changes.

Project Vision

Context Engine is evolving toward:

a local-first ambient AI memory operating system.

The current MVP intentionally focuses on:

• reliable capture
• safe persistence
• private local inference
• offline-first workflows

before expanding into passive intelligence or cloud synchronization.

License

License has not yet been finalized.

Repository Topics

Recommended GitHub topics:

react-native local-ai offline-ai on-device-ai litert whisper ios mobile-ai markdown llm personal-knowledge ai-memory