A (Quantum) Random Access Code ((Q)RAC) is a scheme that encodes $n$ bits into $m$ (qu)bits such that any of the $n$ bits can be recovered with a worst case probability $p>\frac{1}{2}$. Such a code is denoted by the triple $(n,m,p)$... It is known that $n<4^m$ for all QRACs and $n<2^m$ for classical RACs. These bounds are also known to be tight, as explicit constructions exist for $n=4^m-1$ and $n=2^m-1$ for quantum and classical codes respectively. We generalize (Q)RACs to a scheme encoding $n$ $d$-levels into $m$ (qu)-$d$-levels such that any $d$-level can be recovered with the probability for every wrong outcome value being less than $\frac{1}{d}$. We construct explicit solutions for all $n\leq \frac{d^{2m}-1}{d-1}$. For $d=2$, the constructions coincide with those previously known. We show that the (Q)RACs are $d$-parity-oblivious, generalizing ordinary parity-obliviousness. We further investigate optimization of the success probabilities. For $d=2$, we use the measure operators of the previously best known solutions, but improve the encoding states to give a higher success probability. We conjecture that for maximal $(n=4^m-1,m,p)$ QRACs, $p=\frac{1+\frac{1}{(\sqrt{3}+1)^m-1}}{2}$ is possible and show that it is an upper bound for the measure operators that we use. When we compare $(n,m,p_q)$ QRACs with classical $(n,2m,p_c)$ RACs, we can always find $p_q\geq p_c$, but the classical code gives information about every input bit simultaneously, while the QRAC only gives information about a subset. For several different $(n,2,p)$ QRACs, we see the same trade-off, as the best $p$ values are obtained when the number of bits that can be obtained simultaneously is as small as possible. The trade-off is connected to parity-obliviousness, since high certainty information about several bits can be used to calculate probabilities for parities of subsets. (read more)