AI game characters can think locally—and still forget to show up

Can the game keep moving while characters think?

Generative agents use a large language model to produce plans, conversations and memories during play. The engineering problem is that generating a decision can take seconds, while the game must continue updating movement and interactions.

In a November 2026 preprint, Piero Molino and colleagues describe GARP, built inside the Godot game engine. Its Character Behavior System runs with a local, open-weight model containing three billion parameters, without requiring cloud inference. Designers supply characters’ personas and goals; the software combines those inputs with memories, needs and nearby opportunities to choose actions.

The paper reports an implemented system and scenario tests—not evidence that these characters think or feel as people do.

Making a smaller model’s job manageable

Rather than ask the model to consider everything in the world, GARP first filters the available actions. An object’s available action is called an affordance: a bed offers sleeping, for example, while a stove offers cooking.

The filter ranks actions using their relevance to the current goal, distance and whether they belong to the character’s home. It also removes duplicate descriptions. More than 50 candidates become typically three to eight options. In the authors’ action-selection comparison, this reduced response time from about 20 seconds to about eight seconds in the tested configuration.

Model responses must follow structured JSON formats, and the system checks them before execution. Formatting constraints do not guarantee a sensible answer, so failed requests have fallback behavior.

Characters also track six unmet needs: hunger, rest, hygiene, social contact, fun and comfort. These are numerical simulation variables, not measured experiences. Their urgency generally grows with game time and falls through relevant activities.

Four personality axes affect both prompt wording and numerical rules. Extraversion, for instance, changes how quickly social need grows; introversion reduces conversation initiation. The distinction matters because the authors found that personality descriptions alone could change dialogue without reliably changing activities.

Responsive does not mean instantaneous

GARP keeps model requests off the game’s main execution path, reuses cached information and starts some decisions before the current activity ends. Characters can continue their previous activities when new decisions are late.

Its adjustable clock is another important compromise. At the default speed, one real second represents five game-minutes, making a simulated day last about 4.8 real minutes. A 13-second model response therefore consumes roughly an hour of game time. The five-agent birthday scenario runs at one-quarter of that default speed to give conversations more time.

The workload remains substantial: approximately 130 chat-model calls and 80 embedding calls per character per simulated day before the reported reductions. Background execution keeps the game responsive; it does not remove the underlying computation.

The configuration used for updated demonstration captures paired a quantized Llama 3.2 3B Instruct model with an Apple M1 Max. The paper does not establish equivalent performance on other deployment targets, nor a maximum supported population or frame rate across devices. Local operation avoids inference API charges, not hardware or electricity costs.

Remembering is not the same as planning

In an earlier evaluation, the researchers ran nine scenarios three times each. Log-based checks covered active behavior, script or language-model failures, and scenario goals where defined. Twenty-five of 27 runs passed. Shutdown-only callback errors were classified separately.

Both failures occurred in the birthday-party scenario. All five characters reached the host’s house in one run; four arrived in each of the other two. Across a broader set of 11 party runs, including those three, seven achieved full attendance and every run brought at least four characters together.

In every observed party failure, the missing character had remembered the event but omitted it from the daily plan. The breakdown was not navigation: stored information failed to become a scheduled commitment.

What the demonstrations establish

Separate personality demonstrations showed contrasting outcomes, but they were selected successful captures rather than a repeated-run reliability benchmark. Each variant could receive up to three takes; the introverted variant passed on its third.

The researchers did not use human raters to assess believability, isolate each subsystem’s contribution or compare model sizes under matched conditions. Need rates and memory parameters were hand-tuned, and memory drift over simulated weeks remains unmeasured.

GARP demonstrates a locally running architecture that combines generated decisions with explicit game rules. Its party failures expose a narrower, concrete challenge: making remembered commitments reliably survive the next round of planning.