Agentic data science with the Data Agent Kit
cloud.google.com·6h ago
AI Research
arXiv CS.AI·Sep 9·5 min read
MeClear: Cooperative Game-Theoretic Attribution and Risk-Aware Memory Clearance for Long-Horizon LLM Agents
TL;DR
MeClear is a novel framework designed for long-horizon Large Language Model (LLM) agents that enhances memory management by selectively suppressing memories with negative utility. It achieves a significant improvement in task recovery rates, demonstrating a 25.5 percentage point increase over traditional methods.
✦ Why It Matters
Implement MeClear to improve memory management in your LLM applications for better task performance.
Key Takeaways
- Integrate MeClear for enhanced memory management in LLMs.
- Utilize cooperative Shapley attribution for better memory evaluation.
- Aim for a recall rate above 85% in memory systems.
- Consider task-conditioned approaches for memory clearance.
How It Works
MeClear employs a dual approach combining Leave One Out screening and cooperative Shapley attribution. Leave One Out screening evaluates the impact of removing individual memories, while cooperative Shapley attribution distributes utility across multiple pieces of evidence.
This allows MeClear to identify and suppress memories that negatively affect task performance without permanently altering the memory bank.