Tacit Research is a lab studying how models can learn from the work they do.

Every language model stops learning the day its training ends. Everything after that, the codebase it works in, the documents it reads, the corrections it gets, survives only as notes the model writes to itself and reads back the next time. Each task begins with a model that remembers nothing, working from the notes of a previous self.

Our work is on memory that lives between context and weights: representations of a document that are computed once, cost a fraction of the original to carry, and compose with each other. The science here is young and scattered across papers that don’t agree on how to measure it. We think it is the next axis of scaling, and we do it in the open.