A Memory is a short, plain-text note about something worth remembering — a bug's real root cause, a decision you made and why, a preference an agent should keep in mind next time. Each one attaches to whatever it's about: a workspace, a Domain, a repository, a thread, or several other kinds of things.
Agents write these down themselves as they work, whenever they land on something worth keeping — you don't have to ask for it. They're append-only by design: a newer note supersedes an older one just by being more recent, rather than editing history.
The payoff shows up over time, not immediately — a workspace an agent has worked in for a few weeks starts every new conversation a little smarter than the last, because the things it already figured out are still sitting right there waiting to be reused. You can add one yourself too — see Turning a highlight into a memory.