We resolve the question of optimality for a well-studied packetized implementation of random linear network coding, called PNC. In PNC, in contrast to the classical memoryless setting, nodes store received information in memory to later produce coded packets that reflect this information. PNC is known to achieve order optimal stopping times for the many-to-all multicast problem in many settings. We give a reduction that captures exactly how PNC and other network coding protocols use the memory of the nodes. More precisely, we show that any such protocol implementation induces a transformation which maps an execution of the protocol to an instance of the classical memoryless setting. This allows us to prove that, for any (non-adaptive dynamic) network, PNC converges with high probability in optimal time. In other words, it stops at exactly the first time in which in hindsight it was possible to route information from the sources to each receiver individually. Our technique also applies to variants of PNC, in which each node uses only a finite buffer. We show that, even in this setting, PNC stops exactly within the time in which in hindsight it was possible to route packets given the memory constraint, i.e., that the memory used at each node never exceeds its buffer size. This shows that PNC, even without any feedback or explicit memory management, allows to keep minimal buffer sizes while maintaining its capacity achieving performance.