Explain: Don't communicate by sharing memory; share memory by communicating

This famous quote can be a little bit confusing if taken too litterally. Let's break it down to its more basic components, and define them properly:

Don't communicate by sharing memory; share memory by communicating
      ---- 1 ----    ------ 2 -----  ---- 3 -----    ----- 4 -----

1. This means that different threads of execution would be informed of the change of state of other threads by reading memory that would be modified somewhere else. A perfect example of this (although for processes instead of threads) is the POSIX shared memory API: http://man7.org/linux/man-pages/man7/shm_overview.7.html. This technique needs proper synchronization because data races can occur pretty easily.
2. Here this means that there is indeed a portion of memory, physical or virtual, that can be modified from several threads, and read from those, too. There is no explicit notion of ownership, the memory space is equally accessible from all threads.
3. This is entirely different. In Go, sharing memory just like above is possible, and data races can occur pretty easily, so what this actually means is, modifying a variable in a goroutine, be it a simple value like an int or a complicated data structure like a map, and give away ownership by sending the value or a pointer to the value to a different goroutine via the channel mechanism. So ideally, there is no shared space, each goroutine only sees the portion of memory it owns.
4. Communication here simply means that a channel, which is only a queue, allows one goroutine to read from it and so be notified of the ownership of a new portion of memory, while another one sends it and accepts to lose ownership. It is a simple message passing pattern.

In conclusion, what the quote means can be summed up like this:

Don't overengineer inter-thread communication by using shared memory and complicated, error-prone synchronisation primitives, but instead use message-passing between goroutines (green threads) so variables and data can be used in sequence between those.

The use of the word sequence here is noteworthy, because it describes the philosophy that inspired the concept of goroutines and channels: Communicating Sequential Processes.

In essence, yes. Any values that are assigned to variables before the channel send are eligible to be observed after the channel read, since the channel operation imposes an ordering constraint. But it's important to remember the other part of the equation: if you want to guarantee that those values are observed, you have to make sure that no one else can write to those variables in between the write and the read. Obviously using locks would be possible, but pointless at the same time because if you're already combining locks and cross-thread memory modification, what benefit are you getting from channels? You could pass around something as simple as a boolean, as a token allowing exclusive access to the global data, and it would be 100% correct in terms of the memory model guarantees (as long as your code didn't have bugs), it would just probably be a bad design because you would be making things all implicit and action-at-a-distance-y without a good reason; passing the data explicitly is usually going to be clearer and less error-prone.

I don't think so. The gist is instead of protecting one fixed memory address with a lock or other concurrent primitives, you can architect the program in a way that only one stream of execution is allowed to access this memory by design.

The easy way to achieve this is to share the reference to the memory by channels. Once you send the reference over the channel you forget it. In this way only the routine that will consume that channel will have access to it.