Cortex

A private AI assistant that actually remembers you. Everything runs locally on your phone - no cloud, no data sharing, just you and your AI companion.

The Problem

Every AI chat starts from scratch. You introduce yourself, explain your preferences, share context about your life - and then the conversation ends and it's all forgotten. Cloud assistants claim to "remember" but they're really just storing your data on someone else's servers.

Our Solution

Cortex runs entirely on-device using the Cactus SDK. Your conversations, memories, and personal data never leave your phone. It learns who you are over time and genuinely remembers - your name, your job, your preferences, your past conversations.

Memory Architecture (Track 1: Memory Master)

We modeled Cortex's memory after how human brains actually work. Four distinct memory systems work together:

Episodic Memory - Your personal timeline. Every conversation gets stored with timestamps, so Cortex can recall "what we talked about yesterday" or "that idea you mentioned last week." These are full experiences, not just facts.

Semantic Memory - The facts about you. When you mention you're a software engineer, have a dog named Max, or prefer dark mode - Cortex extracts these and stores them separately. The LLM uses tool calling to decide what's worth remembering. These facts persist across conversations and get surfaced when relevant.

Procedural Memory - How you like to communicate. Do you want short answers or detailed explanations? Formal or casual tone? Cortex learns your preferences and adapts. These are the "rules" for how to interact with you specifically.

Working Memory - The active context. Limited to 7±2 items (Miller's Law, just like human short-term memory). When you're mid-conversation, this holds the most relevant episodic memories, semantic facts, and procedural rules. Items compete for these slots based on relevance and recency.

Memory Decay - Memories aren't forever. Just like human memory, unused information fades while frequently accessed memories strengthen. Each memory has an "importance" score that decays over time but gets boosted when retrieved. This keeps context relevant without infinite storage bloat - old, unused memories naturally fade away while important recurring information gets reinforced.

Real Utility Beyond Chat

• Voice memos transcribed and remembered
• Photos analyzed and stored in memory
• Automatic fact extraction from conversations
• Memory explorer UI to see what Cortex knows about you

Why Small Models Work

We deliberately chose the smallest models that get the job done. This isn't a limitation - it's a feature.

• Qwen 3 0.6B for chat (394MB) - Tiny but supports tool calling, which is essential for our memory system
• Nomic Embed for vectors (533MB) - Fast semantic search without cloud APIs
• LFM Vision (420MB) and Whisper Small (464MB) - Multimodal without the bloat

The result: Cortex runs smoothly on mid-range phones, not just flagships. AI should be accessible to everyone, not just people with the latest hardware. These models load in seconds and respond instantly because they're small enough to actually fit in mobile memory.

What Makes It Unique

100% Offline & Private - Works in airplane mode. No internet required after initial model download. Your data physically cannot leave the device.

Edge Capabilities

Feature	Implementation
Offline	All inference runs locally via Cactus SDK
Zero Latency	Streaming responses, no network round-trips
Total Privacy	No telemetry, no cloud, no data leaves device

Tech Stack

Built with Flutter + Cactus SDK. Four models swap in/out of memory as needed since mobile can't keep everything loaded at once.

Purpose	Model	Size
Chat	Qwen 3 0.6B	394 MB
Embeddings	Nomic Embed V2	533 MB
Vision	LFM 2 VL 450M	420 MB
Speech	Whisper Small	464 MB

Project Structure

lib/
├── core/
│   ├── models/           # Memory data structures
│   ├── services/         # Cactus SDK wrapper, memory manager
│   └── tools/            # LLM function calling for memory
├── features/
│   ├── chat/             # Main conversation UI
│   ├── memory/           # Memory explorer views
│   └── onboarding/       # Loading & setup
└── shared/               # Theme, constants, widgets

Pre-built Downloads

Platform	File	Notes
Android	cortex-release.apk	Ready to install
iOS	cortex-release.ipa	Ready to install

Available in builds/ folder and repo root.

Building

# Android
flutter build apk --release

# iOS (requires Apple Developer account)
flutter build ipa --release --export-method=development

Cactus SDK Integration

See CACTUS_SDK.md for technical details on how we use the SDK.

License

MIT