A language you stop using does not vanish. It moves further back. The words are still filed, but the path to them gets longer, and the pause before you speak grows from half a second to two, and two seconds is long enough for the other person to switch to English for you. That is the whole loss, and it happens quietly over months.
The research on this is old and dull and consistent: what you retrieve, you keep. Roediger and Karpicke showed it plainly in 2006: students who were tested on a passage remembered far more of it a week later than students who reread it, even though the rereaders felt more confident. Bahrick, testing people up to fifty years after their school Spanish, found the same shape from the other end: what survived the first few years stopped fading, and the loss was in reaching the words rather than in the words themselves. Recognition is cheap and retrieval is expensive, which is why a vocabulary app can feel productive while the speaking gets harder. Picking the right answer from four options asks almost nothing of the path you are trying to protect. Producing the sentence yourself, with nobody prompting you, asks everything of it.
Ten minutes of conversation does a strange amount of work here. You retrieve dozens of words under mild pressure, in an order nobody chose for you, and you find out immediately which ones are missing, because the sentence stalls. Drills cannot reproduce that, and neither can reading. The pressure is the point, and it has to be low enough that you do it on a Tuesday when you are tired.
What makes it stick is not intensity but the answer on the other side. Someone writes back, remembers what you said last week, and asks about it again, and the words you used come back a second time without a review queue. That is the mechanism we built around: ordinary conversation, often enough to be a habit, with one letter on Sunday that counts what actually happened.
Sources: Roediger and Karpicke, Test-Enhanced Learning, Psychological Science 17(3), 2006. Bahrick, Semantic memory content in permastore, Journal of Experimental Psychology: General 113(1), 1984.
