Virtual Agent Memory Management for Personalized Interactions
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current virtual agent technologies struggle to create deep, personalized interactions with players, as they lack effective mechanisms for updating and refining their behavior and memory based on continuous player feedback, leading to limited emotional bonding and responsiveness.
Innovation Solution
A system that includes a virtual agent memory mechanism, a prompt generation layer, and an interpretation/action layer, utilizing a Large Language Model (LLM) to generate human-like responses, which updates the agent's memory and behavior based on player interactions, feedback, and game state, enabling dynamic and personalized responses.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a virtual agent uses a Large Language Model to generate human-like responses, then the emotional bonding and personalization are improved, but the system complexity and computational resources increase
Solution Approach 1:
The system is divided into distinct functional layers: a prompt generation layer that prepares input for the LLM, the LLM itself for text generation, and an interpretation/action layer that processes outputs. This segmentation allows each component to be optimized independently while managing overall system complexity.
Solution Approach 2:
The prompt generation layer acts as an intermediary between the game state/memory and the LLM, translating complex game states into structured prompts. The interpretation/action layer serves as another intermediary between the LLM's text output and the virtual agent's actions, converting natural language into executable commands.
2Loss of information
If the virtual agent memory stores ever-growing information about player interactions, then the emotional bonding deepens, but the prompt input size limitations are exceeded
Solution Approach 1:
The system extracts only the most relevant information from the ever-growing virtual agent memory into the prompt, rather than including all historical data. This selective extraction allows the prompt to remain within size limitations while still capturing essential context for personalized interactions.
Solution Approach 2:
The prompt generation layer performs preliminary processing of memory information before it reaches the LLM, organizing and filtering data in advance. This preliminary action ensures that only necessary information is included in the final prompt, managing input size constraints effectively.
3Adaptability or versatility
If the virtual agent provides unique, ever-evolving interactions based on continuous learning, then the emotional bonding increases, but the computational processing time increases
Solution Approach 1:
The system updates virtual agent memory and re-evaluates player preferences at periodic intervals rather than continuously processing every interaction in real-time. This periodic action reduces computational overhead while still enabling continuous adaptation over time.
Solution Approach 2:
The interpretation/action layer is customized uniquely to each player, with different processing rules and memory update strategies applied locally based on individual player behavior patterns. This local quality optimization reduces overall processing time by avoiding uniform complex processing for all interactions.
Data Source
AI summary
A method may include: receiving, by a computer program, a user speech or a user action from a user of the computer program, the computer program comprising a virtual agent; identifying a user intent from the user speech or the user action; retrieving saved user-specific memories, static data, and an application state for the computer program; generating a prompt based on the user speech or the user action, the saved user-specific memories, the static data, and the application state; providing the prompt to a text generation module and receiving a suggested action for the virtual agent; converting the suggested action into virtual agent speech and a virtual agent action, wherein the virtual agent outputs the virtual agent speech and takes the virtual agent action; and updating the saved user-specific memories, the static data, and/or the application state with the virtual agent speech and the virtual agent action.