AI Tools & Applications
MemoryLens — AI Semantic Search Engine
Find a movie, book, song or game from a half-remembered fragment.
- top-1 accuracy, full pipeline
- 92%top-1 accuracy, full pipeline
- top-1, retrieval only
- 54%top-1, retrieval only
- benchmark queries
- 48benchmark queries
- categories
- 6categories
// at a glance
The problem and the idea
The problem
Everyone has had a title on the tip of their tongue. Normal search needs keywords you no longer remember, and chatbots often invent titles that don't exist.
The idea
You pick a category and describe what you remember in plain language. MemoryLens searches a real catalog with hybrid retrieval, reranks the candidates and lets an LLM reason over the shortlist. Each match comes with a confidence score and an explanation. When the catalog has no confident match, it names the real title from model knowledge and labels it as such.
// who it's for
Who it's for
- 01
People trying to recall a film, series, book, song, game or actor
- 02
Readers and gamers who remember a scene or a feeling, not a name
- 03
Anyone tired of chatbots confidently inventing titles
// how it works
How it works
1/5 Understand the query
Swipe sideways to see the whole diagram →// screenshots
Screenshots
// key features
Key features
- Six categories: movies, TV series, songs, books, games and actors
- Grounded in a real catalog, so titles are never made up
- A soft "switch category?" hint when a query clearly belongs elsewhere
- Accounts with search history, per-user rate limits and a daily quota
- Local embeddings and reranking; the LLM costs about $0.0003 per search
- Production setup with nginx, Redis caching, security headers and CI
// design decisions
Design decisions
- 01
Every stage was measured on a fixed 48-query benchmark instead of guessing which part helps.
- 02
Retrieval mostly improves depth (recall@5); the LLM layer is what moves the top-1 answer.
- 03
Embeddings and reranking run locally, so only the final reasoning step costs money.
// built with
Built with
- FastAPI
- React
- PostgreSQL
- pgvector
- Redis
- sentence-transformers
- Docker
Run it locally
$ docker compose up --build