Learning What Not to Forget: Long-Horizon Agent Memory from a Few Kilobytes of Learning

The article introduces LRE (Learned Relevance Eviction), a language-model-free scorer designed to optimize memory management in long-running language-model systems by identifying and retaining critical interaction history. LRE operates with a minimal memory footprint of a few kilobytes and demonstrates superior performance in accuracy and efficiency, outperforming traditional eviction policies by completing tasks with 37% fewer calls and reducing peak context size by up to 52%. This approach is significant for AI practitioners as it enhances memory fidelity in LLM agents without incurring additional computational costs, allowing for more effective handling of long-term interactions.