The low-memory game — the smaller build a 1996 machine got

Prerequisite: Audio — TRK / SFX / 11K / SND (what a bank is) and Audio at runtime (how a name finds one).

Titanic shipped with a second, quieter version of itself. Not a different build and not a menu option — a set of branches inside the game's own scripts that fire when the machine reports too little free memory, swapping half the songs out and switching part of the ambience off.

The port has memory to spare, so those branches never ran. This page is what they do, how they were measured, and the one number the play page moves to let a player ask for them anyway.

Reference implementation: GameSession.lowMemory (engine/session.ts) and heapsize (engine/builtins/helpers.ts).

The switch is the game's, not the engine's

TI.EXE has a lowmemory builtin. Nothing in TAOOT calls it. The BOOTFILE defines a function of the same name, which shadows it, and asks a different question:

code lowmemory ()
    global lowmemory

    lowmemory = 6144000
    if heapsize () < lowmemory
        return true
    endif
    return false
endcode

Under 6 MB free — 6144000 bytes — and it was 1996, so that was a real machine rather than a pathological one. heapsize() is the engine probe; everything downstream is script.

This matters for where the feature lives. The port implements none of the behaviour: it answers heapsize() with 4 MB instead of 64 and TAOOT does the rest. The setting is one lie in one builtin.

The five sites that branch on it

script	site	what it does when memory is short
BOOTFILE	setupdecksound	opens decka.11k instead of decka.trk — and the same for deckb, decke, deckf, cargo
BOOTFILE	setupsinksound	opens sink<phase>.11k instead of sink<phase>.trk
EXTRA.CST	setupboatdeck	skips crowdcrickets()
BOOTFILE	openset	setparam(1, 0), setparam(2, 0)
MAP.STG	openstage	stageparam(1, 0), stageparam(2, 0)

lowmemory() is re-read at each of those, and openset runs per room, so the answer is allowed to change mid-game: a room entered after the machine got tight is quieter than the one before it.

The theme swap

    case "b59"
    case "b70"
    case "cafe"
    case "hallb"
        if lowmemory ()
            opentrackfile ("deckb.11k")
        else
            opentrackfile ("deckb.trk")
        endif
        playnewtheme ("deckb.trk")

Note the last line: whichever file was opened, the theme is asked for by its .trk name. That is not a bug and not a fallback — a .11k bank's trackName field says deckb.trk. The file name and the track name inside it are deliberately different, so one playnewtheme serves both branches. See how a name finds its bank; this is one of the 27 banks (of 92) whose two names disagree.

The crowd

setupboatdeck is the only site that isn't a file swap:

    if not lowmemory ()
        crowdcrickets ()
    endif

crowdcrickets() puts five looping party sounds on the boat deck, each pinned to one of the deck's life* stars — soundloop("party1", true) plus makecricket("party1", starxyz("life14", 1), starxyz("life14", 2), 2000, 0, 0), five times, with a different star set per phase. They are crickets: positional ambient sound that pans and fades with where you stand. Skip the call and mission 4's boat deck has no crowd on it at all.

This is the one change you cannot miss, because it is a presence rather than a duration.

.11K is not 11 kHz

The name says sample rate and means nothing of the kind. Measured across every bank that has a .11k twin:

bank	loop chunks	full	short	cut
decka	11 → 6	72.1 s	37.4 s	−48%
deckb	17 → 8	101.9 s	51.3 s	−50%
decke	20 → 10	139.5 s	70.1 s	−50%
deckf	12 → 7	88.2 s	43.3 s	−51%
cargo	11 → 6	74.8 s	37.5 s	−50%
sink0	12 → 6	84.1 s	39.8 s	−53%
sink1	11 → 5	74.8 s	27.1 s	−64%
sink2	18 → 5	83.6 s	34.0 s	−59%
sink3	6 → 6	36.4 s	36.4 s	0%
sink4	12 → 6	69.4 s	35.7 s	−49%
sink5	8 → 6	57.0 s	38.4 s	−33%

Same codec (v40, 8-bit ADPCM), same 22050 Hz — durations are of the concatenated loop chunks as the port assembles a theme. They are the short versions of the songs, roughly half the loop, and the two .trk banks that mix 11025 Hz chunks into their lists (deckb, decke) have .11k twins that don't.

sink3 is the odd one out and worth stating precisely, because it is easy to get wrong: the two files are the same size and differ in 592 bytes, all inside the first 1531 — the container tables, not the audio. Decoded, both give 801792 samples at 22050 Hz with the same hash. Sinking phase 3 sounds identical either way.

Eleven .11k names ship, and the same eleven in all six language editions: cargo, decka, deckb, decke, deckf, sink0–sink5. (Sixteen files per tree — the five deck banks appear on both discs.)

The two params are invisible

setparam and stageparam are otherwise pure scratch: the scripts write 1 and 2 here and no script ever reads them back. So their meaning lives in TI.EXE, and it is not anything on screen.

The setter stores two words:

0x409d12:  mov word ptr [0x489f5c], ax    ; setparam(1, n)
0x409d23:  mov word ptr [0x489f5e], ax    ; setparam(2, n)

and the only code that reads them outside the matching getter is the set-open path at 0x43aa30:

• param 2 (0x489f5e, tested at 0x43aa57) gates a look-ahead: when it is set and there is a next index to reach for, the engine runs an extra load/lock/fetch (0x424000, 0x4222d0, 0x422210) before it needs it. A prefetch.
• param 1 (0x489f5c, tested at 0x43ab3d) picks between 0x438850 and 0x438900. The two functions are otherwise instruction-for-instruction the same; they diverge only after the reference count is decremented, where the param-1-off path calls an extra 0x438b80 when the last reference has dropped. A free-immediately versus keep-cached knob.

Both are memory-pressure knobs, and both are answers to a problem this port does not have: it has its own decoded-ring LRU (ring-cache.ts) and warms neighbouring rings itself. So they stay the scratch words setparam already stored, and everything the setting reaches here is sound.

That is also why the play page names the row for the condition — Low memory — and not for the result. What a small machine got is the game's answer, and it is a different answer in a different port.

Using it

The box is in its own row under the screen; the answer is remembered under taoot.sound.lowmemory. It takes effect at the next openset, i.e. the next room you walk into, not the one you are standing in.

Where it is worth listening:

where	what changes
mission 4, the boat deck	the crowd disappears — the clearest of all
Scotland Road (scot1–scot3), the bridge	139.5 s → 70.1 s, the largest cut
your cabin b59, b70, the café, B-deck hall	101.9 s → 51.3 s, and reachable in the first minutes
F deck: Turkish baths, squash court, ehall, stair2c	88.2 s → 43.3 s
A deck: lounge, smoke room, gym, wireless, grand staircase	72.1 s → 37.4 s
the cargo holds, the bins, crew quarters	74.8 s → 37.5 s
mission 4, sinking phase 1	74.8 s → 27.1 s, the largest proportional cut

And where nothing happens at all, because those setupdecksound cases have no .11k to swap to: the bedsit, deckbd/fore/poop, C and D decks (c73, c78, hallc, recept1c, halld), the engine room and control room, the turbine, and the false smokestacks.

Tested

tests/auto/regression.ts, "the small-memory setting opens the game's own short themes", drives the whole chain rather than any one link: it sets the flag, opens wireless.set (an A-deck room), and checks that the game opened decka.11k instead of decka.trk, that playnewtheme("decka.trk") still resolves to it, and that the theme that comes out is under 60% as long at the same sample rate.