How I Built Memory Concierge: An AI Hotel Concierge That Remembers Guests
I built Memory Concierge, an AI-powered hotel concierge prototype that remembers returning guests and helps hotel staff personalize a stay before the guest even asks.
The project is built around a fictional luxury hotel called The Aurelian Vale.
The idea is simple:
What if hotel service could feel proactive instead of repetitive?
Instead of making guests repeat their preferences every time they return, Memory Concierge helps staff remember, prepare, and adapt the guest experience in real time.
Memory Concierge dashboard for The Aurelian Vale.
Table of Contents
- Why I Built This
- What Memory Concierge Does
- Tech Stack
- How I Used AI
- How The App Works
- Using NVIDIA NIM
- Why Structured JSON Matters
- Real-Time Adaptation
- Adding Voice With ElevenLabs
- What I Learned
- What I Would Build Next
Why I Built This
I wanted to build an AI project that felt like a real workflow, not just another chatbot.
Hotels already collect a lot of useful guest information:
- Room temperature preferences
- Pillow type
- Dietary restrictions
- Favorite drinks
- Morning routines
- Meeting schedules
- Past celebrations
- Arrival details
But that information is often scattered or underused.
A returning guest may still need to repeat the same preferences every time they arrive.
I wanted to explore a different experience:
The hotel already remembers you, and the concierge quietly adapts before you ask.
That became the core idea behind Memory Concierge.
What Memory Concierge Does
Memory Concierge is a staff-facing dashboard.
A staff member can select an arriving guest, and the app generates:
- A personalized welcome message
- Room preparation details
- Dinner or arrival adjustments
- A tomorrow morning brief
- A memory-based concierge note
- Real-time adaptation for events like delayed flights
- Optional voice playback for the welcome message
The main demo guest is Sarah Chen, a returning guest whose flight is delayed by two hours.
The concierge already knows Sarah prefers:
- A high-floor city-view room
- 68°F room temperature
- Firm pillows
- Burgundy Pinot Noir
- Sparkling water
- A yoga mat
- A morning gym routine
When her delayed flight is selected, the app generates an updated stay plan and a calm guest-facing message.
That is the main product idea: proactive hospitality powered by memory.
Tech Stack
I built Memory Concierge with:
- Next.js 14 App Router
- TypeScript
- Tailwind CSS
- Framer Motion
- Lucide React
- NVIDIA NIM
- Llama 3.1 70B Instruct
- ElevenLabs text-to-speech
- Local JSON guest data
- Codex as a development partner
The project is intentionally lightweight. There is no database and no authentication in the prototype. Guest memory is stored locally in data/guests.json so the demo stays simple and reliable.
How I Used AI
AI was used in two different ways.
First, I used Codex while building the project. It helped me scaffold the app, debug TypeScript issues, improve the README, polish the UI, and document the project.
Second, the app itself uses AI at runtime. When a guest is selected, the app sends guest context to NVIDIA's Llama model and asks it to return structured JSON for the dashboard.
So the AI is not just chatting — it is powering a workflow.
How The App Works
The project structure looks like this:
app/
api/
adapt/route.ts
voice/route.ts
welcome/route.ts
globals.css
layout.tsx
page.tsx
components/
AlertBanner.tsx
DashboardHeader.tsx
GuestSelector.tsx
ItineraryCard.tsx
MemoryTimeline.tsx
RoomReadyCard.tsx
TriggerPanel.tsx
VoiceButton.tsx
data/
guests.json
lib/
elevenlabs.ts
json.ts
openai.ts
The main flow is:
- A staff member selects a guest.
- The frontend sends that guest profile to
/api/welcome. - The API route calls the AI model.
- The model returns structured JSON.
- The dashboard renders the welcome, room status, dinner note, and tomorrow brief.
- A staff member can trigger a real-time event.
- The app calls
/api/adapt. - The model returns updated actions and a guest-facing message.
Using NVIDIA NIM
I used the OpenAI JavaScript SDK, but pointed it at NVIDIA's OpenAI-compatible endpoint.
import OpenAI from "openai";
export const NVIDIA_MODEL = "meta/llama-3.1-70b-instruct";
export const openai = new OpenAI({
apiKey: process.env.NVIDIA_API_KEY ?? "missing-nvidia-api-key",
baseURL: "https://integrate.api.nvidia.com/v1"
});
The model used is:
meta/llama-3.1-70b-instruct
I chose this because the project needs clean JSON responses that the frontend can render reliably.
Why Structured JSON Matters
One important design decision was asking the model for structured output instead of free-form text.
The /api/welcome route asks for JSON like this:
{
"welcome_message": "A warm personalized welcome",
"room_status": "Room preparation details",
"dinner_note": "Dinner or arrival adjustment",
"tomorrow_brief": "Morning schedule preview",
"concierge_note": "Personal memory-based note"
}
This matters because each field maps to a specific part of the dashboard. The model is not just generating text — it is generating UI-ready data. That made the app feel much more reliable.
Real-Time Adaptation
The /api/adapt route powers the event buttons in the dashboard.
Example events include:
- Flight delayed 2 hours
- Guest went vegan
- Needs quiet room
- Dinner moved to 9 PM
The model returns:
{
"alert_title": "Flight Delay Handled",
"alert_detail": "The arrival plan has been adjusted around the delayed flight.",
"actions_taken": [
"Updated check-in timing",
"Held room preparation",
"Adjusted dinner reservation"
],
"guest_message": "Your arrival has been adjusted and everything will be ready when you arrive."
}
This is what makes the project feel like an active concierge system instead of a static guest profile page.
Adding Voice With ElevenLabs
I also added optional voice playback.
When the user clicks Play Welcome, the app sends the generated welcome message to /api/voice. That route uses ElevenLabs to generate MP3 audio.
const audioStream = await elevenlabs.textToSpeech.convert(CONCIERGE_VOICE_ID, {
text,
model_id: "eleven_v3",
output_format: "mp3_44100_128",
voice_settings: {
stability: 0.6,
similarity_boost: 0.85,
style: 0.2,
use_speaker_boost: true
}
});
I made voice generation user-triggered because it costs API credits and should only run when the user actually wants to hear the welcome message.
What I Learned
The biggest lesson was that AI products work better when the model has a clear job.
For this project, the model was not asked to be a generic assistant. It was asked to act as a luxury hotel concierge and return specific structured fields. That made the output much easier to trust and display.
I also learned that polish matters a lot in AI demos. Small details like loading states, clear cards, strong README documentation, a screenshot, and a simple story make the project easier to understand.
What I Would Build Next
- A demo GIF showing the full interaction
- More guest profiles
- Database-backed guest memory
- Authentication and staff roles
- Real flight-status integration
- Staff approval before sending messages
- Audit logs for AI-generated actions
- Multilingual concierge responses
- Schema validation for AI responses
- Tests for API routes and UI states
Final Thoughts
Memory Concierge helped me think about AI as a workflow layer.
The most useful AI products may not be the ones that chat the most. They may be the ones that quietly read context, return structured output, and help people act faster.
That was the goal of Memory Concierge:
Make guest memory useful at the exact moment service happens.
Top comments (0)