Learning in a World of Notes (engram.com)

A new study demonstrates how the combination of continual training and effective text memory systems can significantly enhance the intelligence and efficiency of AI agents. By providing models with high-quality notes and employing reinforcement learning, the research reveals that agents can achieve accuracy improvements of 7 to 20 percentage points across various tasks. This innovative approach not only outperforms untrained models and state-of-the-art frontier models but also proves to be cost-effective; training costs represent only 5-30% of building a robust text memory system. Thus, the investment in training pays off quickly, often recouped within a week or less based on user workload. Importantly, the study shows that trained models remain surprisingly robust even as their notes are dynamically updated. They learn generalizable strategies for utilizing notes rather than overfitting to their specific contents, allowing them to maintain their performance despite extensive changes in text data. The findings indicate that training amplifies the benefits of text memory, paving the way for future AI systems that not only interact with notes but actively learn to remember and optimize their use, potentially revolutionizing how AI handles knowledge accumulation and retrieval in ever-evolving environments.