Summary of Changes

Hello @m-mizutani, I'm Gemini Code Assist¹! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!

This pull request introduces an experimental, intelligent memory management system for agents. The core idea is to make agent memories more adaptive and efficient by assigning a quality score to each memory, dynamically re-ranking them during retrieval based on their quality and recency, and automatically pruning outdated or unhelpful ones. This system leverages LLM-based feedback to continuously learn and refine memory scores, ultimately enhancing the agent's performance and decision-making over time.

Highlights

• Agent Memory Scoring: Introduced a quality scoring system for agent memories, ranging from -10 (harmful) to +10 (helpful), to dynamically evaluate their usefulness.
• Adaptive Memory Search: Implemented an adaptive search mechanism that re-ranks memories based on a weighted combination of similarity (50%), quality (30%), and recency (20%), improving the relevance of retrieved memories.
• LLM-based Feedback System: Developed an LLM-based feedback collection system that automatically evaluates the usefulness of memories after each agent execution, updating their quality scores.
• Conservative Memory Pruning: Added conservative pruning rules to automatically delete memories that are critically low-scoring, harmful and unused for extended periods, or moderately harmful and very old, ensuring memory efficiency.
• Configurability and Backward Compatibility: The entire memory scoring and feedback system is highly configurable with default parameters optimized for gradual learning, and it maintains backward compatibility with existing memories.