The story behind a memorial companion built for Google's Gen AI Ideathon
Where It Started 🌇
There are two elders in my neighborhood who've spent almost a decade without their better half.
They still go for the same evening walks. Still sit on the same porch. Still miss them, quietly, in the way grief never really announces itself — it just stays. The universe doesn't work on anyone's schedule, and no app is ever going to change that.
But it made me ask a simpler question:
"If we can't bring someone back, can we at least keep the way they spoke, joked, and loved close enough to talk to?"
Not a replacement. Not a resurrection. Just a companion — one that only ever speaks from real memories, offered by the people who loved that person most.
That question became Smriti.
What Smriti Actually Is
Smriti is a digital memorial companion. A family member records real anecdotes, sayings, and routines about someone they've lost. An AI companion then learns to respond the way that person genuinely would have — grounded strictly in what was shared, never in guesswork.
- 🔗 Try it live: smriti-637076895768.asia-southeast1.run.app
- 💻 Source code: github.com/manish-1614/smriti-memorial-companion
- 🎥 Watch the walkthrough:
Right now, Smriti speaks only through text — and that's on purpose. Before a companion earns the right to speak in someone's voice or show up on a screen, it has to first prove it can be trusted with their words. That's the hard part. So that's the part we solved first.
How It Works (No Jargon, Promise)
- You record a memory — a story, a phrase they always used, a habit only you'd remember.
- Gemini turns it into a "fingerprint." Google's text-embedding-004 model converts that memory into a vector and stores it safely in Cloud Firestore, locked to your account alone.
- You ask something. Your question gets the same fingerprint treatment, and Smriti finds the memories most relevant to what you asked — the way a real conversation naturally pulls up the right story at the right time.
- Gemini replies — grounded, not guessed. Only the memories you actually stored shape the answer. Every reply even shows which memories it drew from, so nothing feels like a black box faking empathy.
And if you bring up something that was never recorded? Smriti doesn't pretend to remember it. It responds warmly, honestly, and simply asks — "Would you like to save this as a memory?" — so next time, it's real.
The One Rule That Matters Most 🔒
Every family's memorial lives in its own locked room. Firebase Authentication at the front door, per-user Firestore rules underneath, server-side token checks on every AI call — no two memorials ever mix, no matter what.
Because a memorial isn't just data. It's someone's private world, handed over by the people who loved them. That deserves more than "pretty careful." It deserves verified, no-exceptions careful.
Grounding before generation. Isolation before intimacy. That's the whole philosophy in six words.
What Comes Next
Smriti's journey doesn't stop at text:
- Today — Text Companion. Grounded, honest, live right now.
- Next — Voice. A reply you don't just read, but hear, shaped by how the memories describe them.
- After That — Virtual Talk Mode. A real-time conversational presence, built on state-of-the-art AI — still following the very first rule: only what was truly remembered, never what was made up. Every future step gets tested against that same promise before it ships. No shortcuts, even for the flashier features. > Concept: Future Vision — not a current feature.
Why This, Why Now
AI is remarkably good at sounding sure of things it doesn't actually know. Everywhere else, that's just annoying. Here, in a companion built for people who are already grieving, it would be unkind.
So Smriti was built around one honest instinct: say only what you know, and be graceful about everything you don't.
That's not a flashy pitch. It's just the right way to build something like this — and maybe that's exactly what makes it worth building. 🌱
Smriti was built for the Google Cloud Gen AI Ideathon (#AccelerateAIwithCloudRun), running on Google Cloud Run, Firebase Authentication, Cloud Firestore, and the Gemini API.
Top comments (0)