The answer is 'no, not always'.
Here's an example: Let $E\to M$ be an oriented Riemannian $3$-plane bundle over $M$, with inner product $g$. Then there is a well-defined bilinear cross-product operation on sections $\times :\Gamma E \times \Gamma E \to \Gamma E$ with the property that, if $X$ and $Y$ are unit-length, orthogonal local sections of $E$ over $V\subset M$, then $X\times Y$ is also unit-length and $\bigl(X,Y,X{\times}Y\bigr)$ is an oriented orthonormal basis of sections of $E$ over $V$.
Now let $\lambda$ be any function on $M$ that is not locally constant and define
$$
[X,Y] = \lambda\ X{\times}Y.
$$
This makes $E$ into an Euclidean bundle of Lie algebras that satisfies your assumptions, but, because $\lambda$ is not locally constant, there is no $g$-compatible connection $\nabla$ on $E$ that makes the Lie algebra structure $[,]$ be $\nabla$-parallel. The reason is that the cross-product will be $\nabla$-parallel as a section $C$ of $E\otimes\Lambda^2(E^\ast)$ (with the connection it inherits from $E$), but the Lie algebra product, which is a section $P$ of $E\otimes\Lambda^2(E^\ast)$, is $\lambda$ times $C$ and so will not be $\nabla$-parallel, which is what your condition requires.
N.B.: Note that the concept of torsion has nothing to do with this problem because $E$ is not assumed to be a subbundle of the tangent bundle, and, in any case, $[X,Y]$ need not be the Lie bracket of vector fields. (I think that this may have misled the (first) commenter above.)