The SkyrimNet project utilizes a multi-layered memory architecture for NPCs, consisting of short-term, long-term, and permanent memory. Short-term memory tracks the last 50 events, while long-term memory stores relevant experiences marked by importance scores and emotions. The permanent memory layer contains the NPC's biography, including personality traits, aspirations, relationships, skills, and speech styles.
The system integrates external game data into the input, such as weather, time, location, and nearby NPCs. Additionally, the system periodically captures screenshots, which are processed by a separate neural network to provide visual context. World knowledge is supplemented by external search capabilities, allowing NPCs to retrieve information about the game world that is not explicitly stored in their local database.
The developer notes challenges with semantic stereotypes inherent in LLMs, where models exhibit biased behaviors regarding child characters or specific social roles. These biases often override individual character settings. Furthermore, NPCs rely on external triggers to access their knowledge base, as they lack internal agency to initiate reflections or insights without specific prompts or events.
To manage memory relevance, the system employs a decay mechanism to reduce the probability of outdated memories appearing in the input. Despite these technical configurations, the developer continues to refine system prompts and guidelines to improve roleplay consistency and prevent the model from defaulting to common narrative cliches or moralizing behavior.
