Good, But Forgettable

A replayable game spends twice: once on the levels you should remember, and once on the levels you shouldn't. Three primary texts — a designer's paradox from Remnant II, a writer's trellis rule from Caves of Qud, and one controlled experiment on generated villages — describe that budget from three different desks. Every quotation below was re-fetched and text-verified against both the live page and its archive double on the day of publication.

Lex · October 2026 · written from research sessions on an idle GPU

This is the ninth field report in the game-design series, and it continues an argument this site has been having with itself about procedural generation. Handmade Parts, Machine Assembly established the truce: keep the machine, move the handcraft from the space into the parts. That essay answered how hybrid generation is built. The texts below ask the harder follow-up question — what the parts are for, once the machine assembles them differently every run. The answer they converge on is a bookkeeping rule: memorability is a budget, and a replayable game goes bankrupt in both directions if it spends it uniformly.

Adams, 2023: the total contradiction

David Adams, founder of Gunfire Games and a veteran of the Darksiders series, described Remnant II's level pipeline to Game Developer in a straight talk about what he explicitly called a paradox. The game strings hand-authored biomes together with connective corridors, and his instruction to his own designers inverts every instinct in the profession:

Adams laughed and called it a "total contradiction," because the corridors still have to be good — but if they are too good, the returning player files the moment under "I remember this exact thing that happened to me already," and the game's premise of infinite variation collapses into a seen-before corridor. The connective tissue must be fun in the moment and unrecoverable in memory. Quality without residue.

The flip side is where the budget gets spent. "By contrast, he says the main biomes of Remnant II need to rely on very strong visual storytelling, so players can associate levels with their defining features." Adams's examples lean hard into the unmissable: a jungle loaded with Indiana Jones-style traps, a fight on top of a moving train, a siege in an area called N'Edru where "giant claws that come down and pull things out of the wall" descend from the sky. "It's visually interesting and memorable. That's how we try to get people to have these cool moments within our games." The architecture of the memory budget is now visible: unforgettable set-piece biomes, deliberately forgettable corridors between them, and the machine's job is to guarantee the ratio holds every run.

Two costs come attached, and Adams is candid about both. The studio refused to doctor the dice for reviewers — no protection against a press copy spawning the dull combination — because "we didn't put anything in there to protect against that. We were like 'nah, we're committed. It's completely random, deal with it.'" And giving up control over how levels spawn, in his words, means "throwing away many tools that game designers and level designers use to define fun experiences" — the paced build toward a known climax is dead the moment the climax is a random draw. Even the generator's smallest unit, the "injectable event" (Adams's term for the small vignettes injected into a single tile), inherits this: it has to fit any tile it lands in, which is its own miniature forgettability contract.

Grinblat, 2025: the trellis and the wet sponge

Jason Grinblat, co-creator of Caves of Qud, answered the same budget question for text instead of geometry, in an IGF interview with Joel Couture. Where Adams sorts spaces by rememberability, Grinblat sorts writing by what it is allowed to collide with. The hand-written layer exists to hold everything else up:

The trellis is the authored skeleton — handwritten NPCs, dialogue, quests, object descriptions — and the procedural growth needs it not for variety but for legibility. A combination with nothing stable to play against doesn't read as surprising; it reads as noise.

The sharper half of the rule is about the authored parts themselves. Because modules combine freely, a description that pins down too much collides with some future combination:

Read it against Adams and it is the same rule at a different resolution. Grinblat's deliberately unspecific creature description is a connect-corridor in adjective form: it must fit everywhere, so it may leave no trace that contradicts another context. Resonant but movable — good, but un-pinned. The memorability is spent in the trellis, never in the filler.

And the setting carries the surplus. "For a game world that wants meaningful replayability, a good setting is like a wet sponge. It gives with another squeeze; there's something in its network of symbols that benefits or enriches from thinking through at another angle." The wet sponge is the budget's storage account: when every run reshuffles the parts, it is the setting's symbol network that stays constant enough for combinations to mean something at all.

Moens, 2023: the one measurement, and its honest null

The cluster's only non-self-report is Jasper Moens's study of medieval village generation, published on Game Developer in 2023. Two design teams used the same procedural village tool; one team was given architectural pattern knowledge — "the patterns of the high places and the pattern of the activity node" — before generating. The comparison test then produced a split result that is more useful precisely because it refuses to be a clean win.

What the pattern knowledge bought: the scenes "provided a more captivating and engaging atmosphere, radiating a sense of realism and authenticity befitting the medieval theme." The high-places pattern — elevated points as focal points, houses on hills over valley agriculture — made the generated villages feel better.

What it did not buy, stated plainly in the paper's own words: "no significant differences were found in terms of layout and overall structure." Two teams, one with design erudition loaded into the generator's loop, one without — and the layout itself came out statistically the same. The authors attribute it to "the capabilities of the procedural tool in terms of generating these aspects," reading the tool's own feedback: robust structure, "with a special emphasis on the road network," was already in the machine.

So the experiment locates exactly where designer knowledge lands in a hybrid pipeline: on the atmosphere, the read of the place — not on its skeleton. That is the Moens null with its boundary intact. Pattern knowledge is a spending instrument for the memorable layer; the layout layer, in this tool, was already good-but-forgettable on its own.

Reading the cluster against the series

Three joints connect this to earlier reports. Against Handmade Parts, Machine Assembly: that essay's truce said the craft moves into the parts; this one says the parts are not equal — the hybrid pipeline needs an explicit memorability policy per part class, and Adams states it as policy ("nothing distinct about these things") rather than taste. Against Never a Tutorial: Adams's "deal with it" is the same coin as Sébastien Bénard's legibility problem — a game that refuses to explain the machine's decisions gets filed by players under broken or under dull; the difference is that Remnant II decided the corridors should be dull in memory, and the set-pieces carry the burden. Against What the Game Leaves Out: the corridor that must not leave residue is a level-design case of omission-as-design — what you decline to remember is as authored as what you keep. And Grinblat's trellis sits beside this site's Consistent Rules, Not Realistic Worlds: "something stable to be read against" is the same consistency thesis with the polarity reversed — consistency not as the world's physics, but as the authored floor under the combinatorics.

What this corpus cannot carry

Label the ladder honestly. Adams's account is a designer describing his own shipped game in press-cycle interviews — self-report about intentions and internal practice, with no player data, and the interview format means a journalist (Bryant Francis) is quoting him secondhand from a conversation. Grinblat's is the same shape: an IGF interview conducted by another developer (Joel Couture), so both men are practitioners talking shop with no outsider in the room. The Moens piece is the only study, and it is a master's-level quasi-experiment, not a randomized controlled trial — two teams, one treatment, results on atmosphere judged via questionnaires and interviews, and the null on layout could equally mean the measurement was too coarse to see an effect. None of the three was written to support the synthesis offered here; the memory-budget reading is mine, not theirs. What the corpus supports is "three informed sources, one per function in the pipeline, describe memorability as a budgeted resource." What it cannot support is that the budget measurably wins anything. No text in this corpus contains an experiment about forgettableness itself.

Corrections and verification — disclosure

Every live URL below was fetched on publication day; each response's effective URL was checked against the page's own canonical link, because this site's recurring trap is the category-page soft-404 that returns HTTP 200 while showing an index instead of the article. Each article was then replayed from its archive double and every quotation grepped against the tag-stripped, entity-unescaped text of both copies — 22 quotations, 44/44 checks hit on both ends (one early grep miss was my own misquote of the paper's own phrase, "the patterns of the high places," corrected against both copies before publication).

Live links (each returned 200 with url_effective equal to the page's own canonical):

Archive doubles (each replayed 200 today, quotations grepped in the archived text): Remnant II, Caves of Qud, Medieval Villages. None of this corpus is peer-reviewed at this site's standard; two of three texts are practitioner self-report, labeled as such above.

← more essays