An editor remembers the shot: a bicycle leaning against a blue wall, just before the rain. The filename is gone from memory, and nobody tagged the scene. This hypothetical example exposes a gap in an archive: having the footage is different from being able to retrieve it when a new project needs it.
Screen Memories, a macOS project shared on Hacker News, approaches that gap with local visual search. Its documentation describes searching images and video scenes in everyday language, with scene results linking to a time inside the original video. Literal searches for visible text and transcribed dialogue are separate modes. The project says inference runs on the Mac after model downloads; we have inspected documentation and code, not independently tested its privacy or retrieval quality.
The mechanism turns selected pictures into numerical representations that can be compared with a description. Similarity supplies candidate matches; a timestamp provides the route back to the footage. The video planner uses shot boundaries and a bounded sampling budget. Despite the project's broad “every frame” headline, that is not exhaustive inspection of every frame. A brief moment can fall outside the sampled pictures.
That limitation is part of the product decision. Denser coverage requires more indexing work. A visually similar result also need not be the intended scene. An editor still has to open the source and judge the surrounding context; a similarity score cannot make that judgment for them.
The opportunity we see is earlier in the creative process than generating another asset. If searching becomes practical at the moment an idea forms, material already paid for and carefully made can become a candidate for reuse. An archive starts participating in new work instead of depending entirely on whoever remembers its contents. That is a proposed consequence, not a measured customer outcome.
For a team with years of footage, a useful first AI product could therefore connect an imperfect human recollection to an inspectable original. The quality test is whether people can recover the material they meant and recognize a wrong match. Producing more content would not solve that retrieval problem.
