A memory access happens when the CPU loads bits from main memory or stores bits to main memory. The CPU will first check to see if it can load the bits from a cache, in which case it will skip the (slow) memory access entirely.
One kind of memory access is a data memory address, when the CPU runs an instruction that explicitly loads or stores data from/to memory. In your examples, the arrows request data memory accesses (or cache accesses). "Data" is a very general term but here it means the bits that the program is operating on, as distinct from the bits that convey the program itself.
Another kind of memory access is an instruction fetch, which happens when the program counter gets a new value as part of running the instructions in the program. It usually just steps forward but a branch instruction can jump to a new location. Either way, no memory access is needed when the instructions are already cached.
In your example, the sequence of register/memory instructions have to load into the CPU before it can execute them. That requires instruction fetches.
(A Harvard architecture CPU has one bus to/from data memory and a second bus to/from instruction memory, so it can fetch from both in parallel. In that case the distinction between an instruction fetch and a data fetch is starker.)
The instructions themselves had to get into memory somehow. An interpreter or loader put them there, treating those instructions as data.
The distinction is whether the data is loading by the explicit request of the program instructions vs. the implicit need to load instructions themselves. The former are easier to cost out. The latter costs depend on the size of the instruction caches, on interruptions taken that might spill this program's instructions from the cache, etc.