Process Heaps ELI5
Everything in Erlang is a Term.
A heap of the process is an array of Terms. A stack is another array of Terms. Stack is located inside the heap. Registers — are array of Terms too! Things on the heap are mostly arrays of Terms, but tagged with a Header Tag (see In Depth: Data Types Memory Layout).
Memory Carriers
Allocation in Erlang happens inside so-called “carriers”. They look like “zone memory”, used by games — big chunks of system heap pre-allocated beforehand. Inside carriers the real allocation is happening. For simplicity, you can imagine simple malloc/re-alloc, that’s how it's done.
This complication is done to fight memory fragmentation and is not important for the understanding. You can see erts/emulator/beam/erl_*alloc.c (many those files, one per strategy of allocation). Emulator has command line flags, which control allocation strategies (see Erlang Documentation flags section).
Allocating Inside Heap
When a process needs some memory, its heap_top is incremented and memory under it is ready to use. Some activities may want to allocate on other process' heaps, for example sending a message will perform a copy to the receiving process.
No bookkeeping is happening inside a process heap. There is no tracking which word belongs where, but it is possible to know what is stored in each memory cell by looking at the tag bits.
Garbage Collection
Garbage collector traces known live values from registers and stack and saves them, then drops everything else.
When a heap comes to its capacity threshold (something like 75%), the process triggers garbage collection. A new bigger heap may be allocated. Scanning generational garbage collection algorithm runs on the heap. The algorithm takes the “ Roots” and moves them to the new heap. Then it scans the remaining parts of source heap and extracts more values, referred by the roots. After the scanning, source heap contains only dead values and algorithm drops it.
"Scanning" means, that garbage collector follows the data Roots from the start to the end, parsing everything it meets. "Generational" means, that algorithm splits data into young and old generation, assuming that data often dies young, and old data is less likely to be freed. Also, the algorithm remembers the old position (mature), where the previous scanning has finished. Any data below it is guaranteed to have not updated since the last scan. This trick reduces amount of scanning and speeds up the algorithm.
In reality the picture is a bit more complicated. There can be one or two heaps with different placement logic applied to them.
Having garbage collector run per process heap allows Erlang to keep collection latency low. Also, it doesn’t pause or affect other processes on other schedulers. This is not a simple topic, but theory is all here: Garbage Collection Book