Reverie - the frame that remembers

Turn the photos that you forget into memories you keep – with an AI frame that understands your photos, you can search them by meaning and hear your memories spoken aloud.

Comment

Inspiration

We take thousands of photos, but nobody ever scrolls back through. Physical photo frames show one picture forever, but the phone galleries are an endless, dumb grid. We asked: what if a frame could actually understand your life – and let you talk to it?

What it does

Reverie is an AI memory frame. The moment a photo comes in, it's understood — not tagged.

  • It understands every photo — Amazon Bedrock Vision writes a warm caption, it reads the scene, mood, people, and color palette.

  • Search by meaning, not filenames — type "sunsets by the water" or "the day we celebrated" and the right memories surface, ranked by relevance.

  • Talk to your memories — ask, "What beach trips do I have?"* and get a grounded answer drawn only from what really happened (RAG).

  • Hears them aloud — each memory is narrated in a warm, genuinely human voice.

  • Reminiscence Mode — a calm, large-type, hands-free, voice-first view built for elderly users and accessibility; press begin, and your life plays back to you.

How we built it

v0-scaffolded Next.js on Vercel → FastAPI → Aurora PostgreSQL + pgvector.

Each upload runs through Bedrock Claude vision, gets embedded by Titan into a 1024-d vector, and is stored in Aurora with a pgvector IVFFlat index for fast cosine search. "Talk to your memories" is retrieval-augmented generation over those vectors.

Challenges we ran into

  • The voice kept getting cut off. The frame auto-advanced on a fixed timer that was shorter than the narration, chopping sentences mid-word. We re-synced the slideshow to wait for the voice to finish before moving on.

  • Serverless broke our filesystem assumptions. On Vercel's read-only runtime, our image directory creation crashed the function on boot. We made storage fall back gracefully to a writable path.

  • pgvector at scale. Building the IVFFlat index blew past the default``; we tuned it and benchmarked the system to 100,000 memories to prove it holds up.

Accomplishments that we're proud of

  • A fully live, deployed product on real Aurora + real Bedrock — not a mockup.
  • Hybrid semantic + keyword retrieval with RRF, benchmarked at 41 ms search over 100k memories.
  • The Memory Atlas — turning an invisible vector space into something beautiful you can explore.
  • Reminiscence Mode — designing for the people who need it most, not just for a demo.
  • A narration voice that actually sounds human, so a memory feels spoken, not synthesized.

What we learned

  • Embeddings have a space, and showing it is powerful. Projecting the vector space into the memory atlas turned an abstract database into an emotional experience.

  • Hybrid retrieval beats pure vector search – Fusing dense and sparse signals with RRF made recall noticeably more reliable.

  • The last mile is the hard mile. A great build and a shipped build are very different things, deployment, env wiring and the access where the real work lived.

What's next for Reverie – the frame that remembers

  • Real hardware — cast Reverie to an actual e-ink or framed display in your home.

  • A knowledge graph of your life — connecting people, places, and events so you can ask, "Show me everyone I was with at the lake."

  • Persistent uploads at scale via S3, and shared family frames so memories can be experienced together.

Built With

Share this project:

Updates