Warning: I am not a differential geometer, so some of the following might not make sense.
Background:
Let $w: (T\Omega)^k \to \mathbb{R}$ be a $k$-tensor on $\Omega$, an open subset of $\mathbb{R}^n$.
We can define the "derivative" of this tensor as follows. $Dw:(T\Omega)^{k+1} \to \mathbb{R}$ is a $k+1$-tensor which enjoys the following approximation property
$$ w(p+v_1)(v_2,v_3,...,v_{k+1}) = w(p)(v_2,...,v_{k+1})+Dw(v_1,v_2,...,v_{k+1})+\textrm{Error} $$
Where the error term is order $|v_1||v_2|...|v_{k+1}|$.
If $f$ is a function, $Df$ is the derivative. Here is an example for $w$ a one form:
$w = Pdx+Qdy$
$Dw = \frac{\partial P}{\partial x} dx \otimes dx+ \frac{\partial Q}{\partial x} dx \otimes dy+\frac{\partial P}{\partial y} dy \otimes dx+\frac{\partial Q}{\partial y} dy \otimes dy$
Notice that if you project $D$ of a one form onto the alternating one forms, you get the exterior derivative.
Call a $k$-tensor $w$ closed if there are open sets $U_i$ covering $\Omega$, and $k-1$ tensors $\eta_i$ on $U_i$ with $w = D\eta_i$ on $U_i$. Call a $k$-tensor $w$ exact if there is a global $k-1$ tensor $\eta$ defined on $\Omega$ with $w = D\eta $.
I am interested in the group of closed $k-$tensors mod exact $k-$tensors.
To generalize this beyond subsets of $\mathbb{R}^n$, it seems clear that we would need a connection, since we need to "evaluate" $w$ at ``the same" tangent vectors, but living at different base points. I know enough about connections to know that they allow this to happen, but not much more unfortunately. I do think there should be a group defined analogously to the one I define above for any manifold with a connection.
My question is simply if this group has been studied before, and if so where I should find information about them in the literature. They would possibly encode more information than the de Rham cohomology groups, including data about the geometry of the manifold instead of just the topology.
Also, any information about the basic properties of these groups (really $\mathbb{R}$ -vector spaces) would be appreciated. Are they finite dimensional for reasonable spaces? Do they depend on only the topology?