(1) Assume that the brain has finite capacity for memory.

(2) Therefore with the infinity of stimulus around us, by (1) there
    isn't enough memory for it all to be stored in full fidelity.

(3) "Live algorithms" -- those that operate while a system is in
    motion -- are incredibly difficult and finicky when compared to
    "offline algorithms."

(4) By (2) and (3), figure that a human's sleep-state is a natural
    opportunity for "compressing" memory while many perceptual inputs
    are attenuated.

(5) Due to (4), postulate that sleep is the time when "short term"
    memory is incorporated into "long term," and extraneous memories
    are selected and discarded.

(6) "Loading" memories from dusty neural recesses to process them as
    in (5) must require "staging" or "working" space. Perhaps regions
    of the brain more usually utilized for perception and immediate
    recall are used to juggle memories around.

(7) Then we can understand a "dream" as perception of such a
    structural "recompression" process (5, 6), noting anecdotally a
    tendancy for dreams to involve both very recent, as well as oddly
    distant memories. And this does not deny "meaning" to dreams --
    they are not random, but in fact would reveal tremendous
    associative structure, concerned as they are with the "indexing"
    and organization of the brain.

(8) But a paranoid consequence of (7) is that once an old memory
    renders through a dream, it may be just shuffling around,
    recompressing, or else on its way forever out, deleted and
    discarded. The "dream" --a second-order perception of the memory
    reprocessing algorithm, can be a "last chance" to keep memories.