In Brief
You can run dozens of believable, emotionally-aware virtual characters on a consumer GPU using a 3 billion-parameter model by combining engine integration, action filtering, caching, and smart scheduling to hide model latency.
ON THIS PAGE
Key Findings
GARP embeds generative agents inside a commercial game engine and runs the full planning, memory, emotion, and relationship stack on local hardware. A layered optimization approach—budgeting inference time with a simulation clock, reducing model calls via affordance filtering and caching, and performing decisions slightly ahead of need—makes real-time local operation practical. The system also adds emotion contagion, graded personality, and explicit familiarity/trust estimates so behavior feels coherent across many stochastic model calls. Demonstrations ran roughly 20 agents with most scenario checks passing, though some coordination failures remain (e.g., party attendance). tool-use pattern
Not sure where to start?Get personalized recommendations
Data Highlights
1≈130 chat-model calls per agent per simulated day (before filtering and caching)
2A 3 billion-parameter model takes about 13 seconds per generation on a consumer GPU
325 of 27 automated scenario checks passed; the town demo ran around 20 agents in the tested build
What This Means
Game developers and simulation engineers who want private, on-device non-player characters can use these patterns to get richer behavior without cloud costs. Product and platform leads evaluating agent reliability should look at the budgeted clock, caching, and affordance filtering as practical levers to make small models feel responsive. Researchers building multi-agent social systems can reuse the emotion, relationship, and memory consolidation components to study trust and interaction dynamics. Autonomous Agent
Key Figures

Fig 3: Fig. 4: Ludwig’s Plan panel shows the daily schedule and the activities the character can decompose and revise as the day unfolds.

Fig 4: Fig. 5: Memories tab of the agent panel at midday: seeded background thoughts (one of them an authored rule, whose condition becomes the card’s header), the night’s sleep as an event, and the summary of a morning conversation titled with the partner’s name, each with its encoding time and location tag and filterable by type (event, thought, chat).

Fig 5: Fig. 6: Lena’s Emotions panel after Dara offers to support her at her recital. The panel shows her current emotion and its intensity, an explanation of her reassurance, and the change from negative to positive valence. This is an illustrative run of the stage-fright scenario.

Fig 6: Fig. 7: Two illustrative free-evening captures. Top: the Persona panel shows the higher (left) and lower (right) sociability settings. Bottom: Mika talks with Ren in one run and watches TV at home in the other. The two frames show different moments in the evening. These captures illustrate the behavior contrast; Table V reports outcomes from the separate trait-controlled evaluation.
Ready to evaluate your AI agents?
Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.
Learn MoreConsiderations
Believability was not validated with human raters, so subjective impressions may differ from the log-based checks. The demo shows about 20 agents at one operating point; maximum population, cross-device frame rates, and multi-GPU scaling were not measured. Several parameters (need decay, consolidation) are hand-tuned and some coordination failures (e.g., guests forgetting to add invitations to plans) still occur. State Inconsistency
Methodology & More
GARP integrates a full generative-agent architecture into the Godot game engine with a native Rust backend for fast vector memory and model routing. Characters have authored personas and graded temperament sliders, and they run three-tier planning (daily, hourly, minute-level) driven by a 3 billion-parameter model. The system augments planning with needs, emotional appraisal (including contagion), and pairwise relationship tracking (familiarity, trust, sentiment). Memory uses a retrieval scheme weighted by recency, importance, relevance, and emotion salience plus a consolidation process inspired by human forgetting curves. Making local inference work required a layered set of engineering choices: a configurable simulation clock that defines how much model work fits into simulated time; proactive decision-making so thinking finishes before actions end; affordance filtering to shrink the action space from 50+ candidates to a handful; multi-tier caching and prompt compression; schema-constrained outputs with fallbacks to reduce parsing failures; and asynchronous, non-blocking model calls so the engine continues to run while agents think. The result is a privacy-friendly, iterative-friendly pipeline that ran roughly 20 agents with 25 of 27 automated scenario checks passing. The approach is practical for developers who want on-device agents, but human evaluation, scaling studies, and automated tuning of parameters remain future work. Event-Driven Agent Pattern Chain of Thought Pattern
Need expert guidance?We can help implement this
Credibility Assessment:
Authors have moderate h-indexes (Piero Molino 17, Federico Bianchi 13) indicating established but not top-tier researchers; arXiv venue and no top-institution signal -> solid/recognized.