"Congruence" in the context of lambda calculus is usually "alpha-congruence". "alpha-congruent" means "differring only if at all in the names of bound variables".
For instance, $\lambda x. x \equiv_\alpha \lambda y.y$ and $\lambda xy.x(xy) \equiv_\alpha \lambda yx.y(yx)$ (bound variables $x$ and $y$ were swapped), but $\lambda x.x \not \equiv_\alpha \lambda x.xy$ (different term structure).
Formally,
$P[\lambda x.M] \equiv_{1\alpha} P[\lambda y.M[y/x]]$
where $P[\lambda x.M]$ means that $\lambda x.M$ is a subterm in $P$, $y$ is a new variable that does not occur free in $M$, and $M[y/x]$ is the result of substituting every occurrence of $x$ with $y$ in $M$;
$P \equiv_\alpha Q \text{ iff } P = P_1 \equiv_{1\alpha} \ldots \equiv_{1\alpha} P_n = Q$
i.e. $P$ is alpha-congruent to $Q$ if there is a chain of bound variable renamings from $P$ to $Q$.
The idea is that alpha-congruent terms "mean" the same, because the choice of names for bound variables does not affect a term's behavior in abstraction and application.