There's 2 ways to remember the sign convention:
If you're trading an exchange-listed spread, then the convention is that going long on the spread A-B implies buying A and selling B. Vice versa, shorting the spread implies selling A and buying B.
If you're trading a synthetically-constructed spread, then this means that you're trading the residual, i.e. the difference between the observed $y_t$ and the $\hat{y}_t$ predicted by your regression model.
The simplest example is a pair trade where you're regressing a series $y_t$ against another series $x_t$. You may assume that there exists a linear relationship between the series and a normally distributed error term $\epsilon_t \sim \mathcal{N}$ such that $\epsilon_t = y_t - \hat{y_t}= y_t -\beta x_t -\alpha $. $\alpha,\beta \in \mathbb{R}$ are parameters that you estimate from past data, e.g. with ordinary least squares.
Often, you'd also assume $\alpha$ falls off at $x_t=0$. Then "buying the spread" implies having positive delta to $\epsilon_t$ which means buying 1 unit of the product with series $y_t$ and selling $\beta $ units of the product with series $x_t$.
You don't even need to remember what it means to "buy a spread" in this case, because the intuition behind your trade is simply that if the observed value $y_t$ is less than the predicted value $\beta x_t$, then you would buy the product with series $y_t$ and sell $\beta$ units of the product with series $x_t$, since the observed value and your prediction should eventually converge somewhere. You just need to remember which variable you used as the predictor $x_t$ and the dependent variable $y_t$ when fitting your model.