Virtual Character Memory Retrieval for Natural Dialogue
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Virtual characters in interactive media experiences, such as video games and chatbots, often rely on scripted dialogue trees, which limit their conversational abilities and fail to incorporate nuances like personality and backstory, leading to robotic interactions and hallucinations in responses.
Innovation Solution
A technique that matches user messages to virtual character memories based on multiple factors, using a machine learning model to generate responses that align with the character's persona and experiences, adapting Retrieval-Augmented Generation techniques to provide more dynamic and natural interactions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If scripted dialogue trees and flows are used to control virtual character interactions, then the interactions follow a predetermined structure, but the virtual character cannot improvise and feels robotic
Solution Approach 1:
The patent segments the character's knowledge into distinct memory types (episodic memories, semantic memories, procedural memories) and uses separate retrieval mechanisms for each type. This allows the system to maintain structured character consistency while enabling flexible, context-appropriate responses by selectively accessing different memory segments based on conversational needs.
Solution Approach 2:
The patent introduces a retrieval-augmented generation system as an intermediary between the user input and the virtual character's response. This intermediary retrieves relevant memories and grounds the generative model's output in the character's established persona and backstory, thereby maintaining consistency while enabling natural, improvised conversation.
2Adaptability or versatility
If generative models are used to enable natural conversations, then the conversations become dynamic and engaging, but the model hallucinates output that is incorrect or not in line with the character's personality
Solution Approach 1:
The patent performs preliminary retrieval of relevant character memories, persona attributes, and backstory information before generating the virtual character's response. By preparing and providing this grounded context in advance to the generative model, the system ensures that the model's natural and dynamic responses remain consistent with the character's established personality and avoid hallucinations.
3Measurement precision
If conventional RAG-based approaches are used to control LLM output, then factual knowledge is supplemented, but nuances of personality and experiences are not fully incorporated
Solution Approach 1:
The patent merges multiple memory types (episodic, semantic, procedural) and combines them with persona attributes and backstory information into a unified retrieval-augmented generation system. This integration ensures that both factual accuracy and nuanced personality traits are simultaneously incorporated into the virtual character's responses.
Data Source
AI summary
One embodiment of the present invention sets forth a technique for synthesizing an interaction with a virtual character. The technique includes matching a first message from a user to a first set of memories associated with the virtual character and determining at least a portion of the first set of memories based on a plurality of factors associated with the first set of memories. The technique also includes inputting a first prompt that includes (i) one or more instructions associated with the virtual character, (ii) the at least a portion of the first set of memories, and (iii) the first message into a machine learning model. The technique further includes generating, via execution of the machine learning model based on the first prompt, a first response by the virtual character to the first message, and causing the first response to be outputted to the user.