0.Demo

Demo for The Little Price

1.Paper

MemoryRepository for AI NPC

2.Guide for read-MemoryRepository for AI NPC

This guide describes the content of each chapter.

Architecture of MemoryRepository

• Introduces the structure of the MemoryRepository,which include three part:Memory Room, Memory Interaction, and Memory Renewal Mechanism.
• Provides a detailed explanation of the Memory Room, Memory Interaction, and Memory Renewal Mechanism.
• Includes figures:
  • FIGURE 1. MemoryRepository Structure.
• Includes tables:
  • TABLE 1. Notation used for modeling and scheduling.

Scheduling Policy of MemoryRepository

• Describes the execution process of the MemoryRepository.
• Explains the scheduling policy of the MemoryRepository, detailing how its three major modules are interconnected and operate.
• Includes figures:
  • FIGURE 2. MemoryRepository Processing.
• Includes algorithm flow:
  • Algorithm 1: Memory Repository Processing.

Example: An AI NPC Game Powered by MemoryRepository

• Details the process of building the virtual town, Example: StarUniverse, using MemoryRepository.
  • Fine-tuned an LLM using open-source NPC-related game data.
  • Integrated the MemoryRepository.
• Describes the experiments conducted in the virtual town:
  • Comparison of NPC dialogues powered by an LLM with and without the MemoryRepository.
  • Human-Like comparison results (Case Study, a qualitative comparison image, no quantitative data).
  • Long-Term Interaction comparison results (Case Study, a qualitative comparison image, no quantitative data).
• Includes figures:
  • FIGURE 3. StarUniverse.
  • FIGURE 4. StarUniverse.
  • FIGURE 5. Prompt Structure of NPC in MemoryRepository.
  • FIGURE 6. Long-Term Comparable.
  • FIGURE 7. Human-Like Comparable.

Experiments and Analysis

• Describes the experimental setup:
  • Platform used.
  • Datasets:
    • Training set.
    • Validation set.
    • Test set.
  • Evaluation metrics:
    • Memory retrieval accuracy.
    • Response correctness.
    • Contextual coherence.
    • Human-Like.
• Explains the experiment:
  • Experimental steps:
    • Data generation.
    • Evaluation methods for experimental results.
  • Includes figures:
    • FIGURE 8. StreamLit-Constructed Local Website for Experimental.
    • FIGURE 9. A segment of these historical records.
  • Includes tables:
    • TABLE 2. Platform Specifications.
  • Experimental results:
    • 1. MemoryRepository Performance and Comparison
      • Includes figures:
        • Figure 10(a) (b) (c) (d).
      • Analysis of figures:
        • (a) shows that retrieval performance slightly decreased with the MemoryRepository.
        • (b) shows that the MemoryRepository did not significantly improve correctness.
        • (c) shows that GPT-3.5 and GPT-4 performed well even without the MemoryRepository.
          • The integration of the MemoryRepository did not significantly enhance coherence for GPT-3.5 and GPT-4.
          • Without the MemoryRepository, ChatGLM performed poorly.
          • With the MemoryRepository, there was a significant improvement in coherence for ChatGLM.
        • (d) shows that the MemoryRepository significantly improved Human-Like performance across all models.
        • Explanation:
          • The MemoryRepository's long-term memory mechanism stores information in the Long-term MemoryRoom through summarization and forgetting processes, discarding less relevant information. This slightly decreases retrieval accuracy but aligns with the goal of focusing on Human-Like and Long-Term Interaction.
          • When a model has coherence issues, the MemoryRepository can provide a significant improvement.
          • If a model already has strong coherence performance, the MemoryRepository cannot offer much improvement.
          • This indicates that the MemoryRepository's memory system endows models without inherent memory mechanisms with both short-term and long-term memory, significantly enhancing coherence.
          • The MemoryRepository provides a notable improvement in Human-Like performance for all models.
          • This is because most models lack a Human-Like memory design, but the MemoryRepository, through its short-term and long-term memory system along with summarization and forgetting mechanisms, makes model dialogues more human-like, thereby significantly enhancing Human-Like performance.
    • 2. Performance Comparison of MemoryRepository and MemGPT
      • Comparison between MemoryRepository and MemGPT using the four evaluation metrics: Retrieval, Correctness, Coherence, and Human-Like.
        • MemGPT slightly outperforms MemoryRepository in Retrieval.
        • MemoryRepository significantly outperforms MemGPT in Human-Like.
        • Correctness and Coherence are similar for both.
        • This shows that the MemoryRepository’s memory mechanism not only enhances long-term interaction but also significantly improves Human-Like capabilities.
      • Includes tables:
        • TABLE 3. Comparison of MemGPT and MemoryRepository-embedded GPT-3.5 Performance in the 100th Round of Dialogue.
    • 3. Performance Comparison of MemoryRepository-Enhanced LLMs Across Diverse Inference Parameters
      • Includes figures:
        • Figure 11(a) (b) (c) (d).
      • Shows that the MemoryRepository exhibits good diversity and robustness, maintaining its functionality even when model parameters are adjusted.
        • It can improve Long-Term Interaction and Human-Like performance.

3.Code - MemoryRepository_ChatGPT

Some code and data related to MemoryRepository are available, but a full codebase is not provided. Only partial implementation and key snippets necessary for understanding and experimenting with the MemoryRepository framework are included.