Problem
Browser history was chronological, while people remembered the meaning of a page.
Constraints
Content extraction, embeddings and query retrieval needed a consistent semantic representation.
What I did
Browser history is chronological; human recall is semantic.
Replay was a Chrome extension that treated your history as a personal search index. It pulled content from the pages you visited, generated embeddings, stored the vectors in Supabase, and ranked history by cosine similarity against a natural-language query.
Instead of remembering a title or a date, you could just describe the page: an article about vector search, a webhook example in some docs, the command that fixed a build.
Product surface
Ctrl-Space opened a small spotlight-style overlay over the current page. The extension stayed invisible until recall failed — the whole interface was tuned for retrieval, not bookmark management.
Engineering surface
It combined page extraction, embedding generation, vector storage, authentication, search, and a React overlay that had to feel fast enough to become a habit. The design decision that mattered was keeping semantic identity consistent across the ingestion and query paths.
Replay shaped how I think about AI products: the model was a dependency, but retrieval was the product.
Result
A browser-history extension that retrieved visited pages from a natural-language description.
Tradeoffs
Semantic retrieval adds an ingestion and vector-storage pipeline; the interface can stay small because the complexity lives in that pipeline.