Games solved character memory decades ago, at the level a game can afford: flags in the save file, dialogue trees gated on those flags, reputation as a number. Bolting a model on top does not change the economics, because a cache per NPC was never shippable. A bounded field gives each character a per-character budget you set, holding the playthrough as it was actually played.
Games have solved character memory for decades at the level a game can afford. Quest state is a set of flags in the save file, dialogue is a tree gated on those flags, and reputation is an integer. It is cheap, deterministic, testable, and it ships. What it is not is memory of what actually happened.
Adding a language model on top does not change the economics. The model gets the flags injected as a prompt preamble, perhaps a retrieved line or two from a lore database, and a context that is cleared when the player quits. So the character can tell you that you finished the quest, because a flag says so, and cannot remember how you finished it, who you spared, or what you said while you were doing it. Multiply a growing cache by every character in a populated world and the reason is obvious: persistent memory has never had anywhere to live at a price a studio can ship. And the character itself — its persona, its lore, the way it speaks — is re-injected as a prompt preamble on every turn, versioned nowhere and drifting whenever the prompt does.
SCOPE: schematic. The measured per-character footprint curve comes out of the pilot, for your title.
A full playthrough of a long game is millions of tokens of dialogue, choices, and events. The question is not whether one character can hold that: it is what holding it costs once the world is populated.
SCOPE: measured on a single unbroken session of 1,900,000 tokens across 9,359 exchanges, 14.5× the base model's trained context, with the resident memory flat from 50,000 tokens through to the final turn and total footprint ~1,370× at 1.9M tokens under a standard cache. The ratio grows with length and is only meaningful at a stated depth. Reproduced on 3 independent runs. That is one character. Aggregate footprint at your intended character count comes out of the pilot.
Characters that remember how the quest went, not just a flag saying it is complete.
The character ships as a lens. Persona, voice, and lore load as one versioned artifact instead of a prompt preamble re-pasted every turn — so the same NPC is the same NPC across a patch, a sequel, or a spin-off.
A per-character memory budget you set, which is what made persistent NPC memory unshippable before.
Continuity you can test. Our published run retains every one of its 9,359 exchanges from its inputs, so a playthrough can be re-run and checked rather than reproduced by hand.
The history stops hitting a ceiling. Before we bounded the read, our own runs died at roughly 285,000 tokens: output collapsed to about a dozen tokens a turn and stopped. A long playthrough passes that. Nothing like it appears at any depth we have run since.
SCOPE: per-character footprint depends on your budget and the character count you intend to ship. Published results measure one field; aggregate behavior across a populated world is what a paid pilot returns for your title. Our depth results were produced on a deliberately small research model, chosen to stress the memory rather than flatter the output. A larger model abuses it less. Treat them as a floor.
The persona, voice, and lore loaded as a versioned lens — the character held apart from the playthrough it reacts to, so it stays itself across sessions and updates without re-tuning a prompt. What a pilot returns is how consistently that lens holds under a full playthrough, including players who spend the session trying to talk the character out of being the character.
Everything the player did with this character carried as one resident structure, so the world remembers how the quest went rather than only that it completed. In our deepest run a thread raised in the opening hour was picked back up eight times across the session, the last ~1.6 million tokens later, and was never once re-introduced by name.
Moment-to-moment dialogue kept word for word over a character block the field cannot overwrite, so the scene in front of the player stays sharp without dropping the history behind it.
How well the character stays in persona and voice across a full playthrough, including under players actively trying to break character.
A per-NPC footprint curve at your intended character count, so a populated world stays inside your memory budget.
Recall on how a quest was completed, not just whether a flag was set, across sessions that currently reset.
Continuity measured over a full playthrough, at each budget setting you choose.
Aggregate footprint across your intended character count, since one field is a number and a populated world is that number multiplied.
SCOPE: this page states direction, not magnitude. Measured numbers for your workload come out of the pilot, with the configuration that produced them.
Bounded memory on its own is not hard. A rolling window reaches the same footprint number tomorrow by dropping the oldest exchanges, and in a game, the oldest exchanges are act one. What is hard is holding the footprint flat while the whole playthrough stays available, and showing the model is still reading the start of it at depth. That is the result we publish, reproduced across independent runs, and it is what a pilot reproduces for your title.
The field is held to a budget you fix before you ship. It runs on the model the title already uses, with no retraining and no recalibration. Patent pending; the implementation stays private.
A scoped four-week pilot: you bring one workload and a memory budget, we return a measured comparison against your current cache: footprint, recall, and the configuration that produced both.