ActiveMem: Distributed Active Memory for Long-Horizon LLM Reasoning

ActiveMem is a newly proposed framework that introduces a distributed active memory system for large language models (LLMs) to enhance long-horizon reasoning capabilities. It features a decoupled architecture where a high-level Planner uses distilled semantic gists for reasoning, while a parallel lightweight memory system accumulates these gists, leading to state-of-the-art accuracy on benchmarks like BrowseComp-Plus and GAIA with reduced computational overhead. This approach addresses the limitations of centralized memory systems, enabling practitioners to build more efficient LLMs capable of handling complex reasoning tasks without context overload.