Just a hunch, but do you find that you perform substantially better if you are an environment where you can talk to yourself out loud as you work through a math or physics problem? If so, this is very much indicative of the phenomenon I'm describing, because what's happening is you're recruiting your audio memory as a compensatory short-term workspace, being able to loop some 2 or 3 additional elements of information in this extra area essentially acts as an extension of your working store. And guess what behavior students are typically prohibited or highly discouraged from doing during an exam?

I further suspect that this is also the mechanism responsible for some people's apparent tendency to do all the talking in a conversation, rather than allow a balanced share of time between conversants. The quick decay of short-term information requires constant refreshing, and talking is one means of augmenting this.