Repo memory is the layer of knowledge that should outlive any single work session: why a pattern was chosen, what must always stay true, what has burned the team before. In most teams this memory lives only in people's heads and in pull-request threads that are never read again. When an engineer leaves or an AI agent starts a fresh session, that memory is gone, and the next person re-derives it — or re-breaks something because they couldn't see it.
The shift that makes repo memory urgent is AI-assisted coding. An agent writes code competently but starts every session with no history; it cannot know the decision behind a piece of code unless that decision was captured somewhere it can read. Without a memory layer, every agent diff quietly buries the reasoning behind it, and the next agent inherits code with no explanation.
Repo memory turns those one-off discoveries into a persistent, shared asset. The loop is simple: after a session you persist what you learned, and before you touch an area again you pull what is already known. The value compounds with the count — a handful of well-anchored notes and every future session, human or agent, starts with the context instead of a blank slate.