Here's a counterexample.

Define $S=\bigoplus_p C_p$, where $C_p$ is cyclic of order $p$ and $p$ ranges over prime. Fix an integer $n\ge 1$. Define $P=\prod_p C_p$, and consider a maximal $\mathbf{Z}$-free family in $P^n/S^n$, generating a free abelian group (of continuum rank) $T/S^n$. Fix a prime $q$ and let $U/S^n$ be (*) a subgroup of $P^n/S^n$ isomorphic to $\mathbf{Z}_q$ (the group of $q$-adic numbers).

Let's check that for every direct decomposition $U=V\oplus W$, either $V$ or $W$ is finite. Since $T(U)\simeq S^n$, this implies that $U$ is a counterexample to your question for $n\ge 3$.

First observe that $\mathbf{Z}_q$ is indecomposable (i.e. has no nontrivial decomposition as direct product, easy exercise (**)). Since $U/T(U)\simeq\mathbf{Z}_q$, it follows that either $V$ or $W$, say $W$, is torsion.

Write $T(V)=\bigoplus A_p$ and $W=\bigoplus B_p$, so that $A_p\oplus B_p=C_p^n$ for all $p$.

We have $P^n=(\prod A_p)\times (\prod B_p)$. So $P^n/T(V)=(\prod A_p/\bigoplus A_p)\times \prod B_p$. Since there is no nonzero homomorphism $\mathbf{Z}_q\to C_p$ for $p\neq q$ and since the subgroup $V/T(V)$ of $P^n/T(V)$ is isomorphic to $\mathbf{Z}_q$, the projection of $V/T(V)$ into $\prod B_p$ is contained in $B_q$. Thus in the above decomposition, $V\subset (\prod A_p)\times B_q$ and $W\subset (\bigoplus A_p)\times (\bigoplus B_p)$. So $U\subset (\prod A_p)\times (\bigoplus B_p)$. Since $P^n/U$ is torsion (because it is a quotient of $P^n/T$ which is torsion), it follows that $\prod B_p/\bigoplus B_p$ is torsion. This means that $B_p=0$ for large $p$. It follows that $W$ is finite.

Here are the two easy verifications:

(*): pick in $\mathbf{Z}_q$ a maximal $\mathbf{Z}$-free family, generating an abelian free subgroup $Z$; since $P^n/S^n$ is a rational vector space, an isomorphism $Z\to T/S^n$ (it exists since both groups are free abelian of continuum rank) can be extended to an injective homomorphism $\mathbf{Z}_q\to P^n/S^n$.

(**): $\mathbf{Z}_q$ is $p$-divisible for all $p\neq q$ and $\mathbf{Z}_q/q\mathbf{Z}_q$ has order $q$, so for every direct decomposition of $\mathbf{Z}_q$, at least one factor is divisible. But $\mathbf{Z}_q$ does not contain any copy of $\mathbf{Q}$.

isfinitely generated, right? Or is the idea that you want to show any such group can arise as the $K$-points of an ab. var. for some possibly infinite extension $K$? In that case, you may assume $G[tors]$ is finite, basically since $\mathbb{Q}_p/\mathbb{Z}_p$ is divisible. – Kevin Ventullo Jun 29 '13 at 4:03anyfield $K$. (And it is not enough to assume that the torsion subgroup is finite. In fact that case is significantly easier than the general case, because the torsion subgroup is necessarily a direct factor.) – Pete L. Clark Jun 29 '13 at 4:29