A voice-first AI companion for dementia care with transparent patient consent controls.

Built for Cal Hacks 12.0 (Phase 11.3) • October 2025

🌐 Live Production: getmemora.xyz 🚀 Deployed on Vercel with automatic deployments from main branch

Memora is a full-stack dementia care platform with:

• Real voice interaction via Deepgram STT/TTS
• Claude Haiku 4.5 AI conversations
• Letta 3-tier memory with ChromaDB archival storage
• Supabase PostgreSQL backend (13 tables, 18 API endpoints)
• Clerk authentication with role-based access

Patient Interface (/patient): Voice-first companion, NO cognitive scores shown. Patient controls what caregivers see via privacy toggles.

Caregiver Dashboard (/dashboard.html): Monitoring analytics with consent notifications. All AI decisions show transparent explanations. Be honest about surveillance, don't euphemize it.

Frontend: Next.js 16 • React 19 • TypeScript • Tailwind CSS • Radix UI Backend: Supabase Cloud PostgreSQL • Prisma ORM • Clerk Auth • 18 API endpoints AI: Claude Haiku 4.5 (responses) • Letta Cloud (memory + ChromaDB archival) • Deepgram (voice) • Groq Kimi K2 (behavioral insights) Testing: Vitest with 37+ tests (real APIs, no mocks)

# Prerequisites: Node.js 18+

git clone https://github.com/yourusername/memora.git
cd memora
npm install
cp .env.example .env            # Fill in cloud API keys (Supabase, Clerk, Anthropic, Deepgram, Letta)
npm run dev                     # Start dev server
# Open http://localhost:3000

# Run tests (37+ tests, real cloud APIs)
npx vitest run

All services are cloud-hosted:

• Production: getmemora.xyz (Vercel, auto-deploys from main)
• Database: Supabase Cloud PostgreSQL (no local DB)
• Memory: Letta Cloud (includes ChromaDB for archival storage)
• Auth: Clerk
• AI: Anthropic Claude, Deepgram, Groq

Key Routes (all accessible at getmemora.xyz):

• / - Landing page
• /patient - Voice interface (React)
• /dashboard.html - Caregiver dashboard (static HTML with client-side routing)
• /sign-in, /sign-up, /onboarding - Auth flow

Hybrid System:

• Static HTML: Landing + caregiver dashboard (public/*.html with client-side JS routing)
• Next.js: Patient voice interface (/patient), auth pages, API routes
• Dashboard uses scale(0.8) transform → all full-height elements must use 125vh

3-Tier Memory (Letta + ChromaDB):

• Core: Agent persona, patient profile, current context
• Archival: Historical conversations with semantic search (ChromaDB embeddings)
• Alerts: Wandering detection, activity patterns

Design: WCAG AAA compliant (7:1+ contrast), 44px+ touch targets, voice-first interaction

See CLAUDE.md for detailed architecture documentation.

• ✅ Voice interface: Fully functional (Deepgram STT/TTS, Claude responses, Letta memory)
• ✅ Groq Kimi K2: Integrated for behavioral analysis (280k context window)
• ✅ Supabase Cloud: 13 tables, 18 API endpoints, all connected
• ⚠️ Dashboard: Uses hardcoded demo data (API calls exist but don't update DOM yet)
• ⚠️ LiveKit: Token endpoint ready but voice UI doesn't use it yet
• Dashboard pages: Only Overview and Voice Chat have real content; Timeline, Insights, Settings are placeholders
• Mock data: data/mock-data.ts still used in dashboard.html (NOT in voice interface or APIs)

MIT License

Built with care at Cal Hacks 12.0