GeoForge Enhances Earth Observation AI Agents with Self-Evolving Knowledge

A new framework named GeoForge has been introduced to enhance the capabilities of Earth observation (EO) agents, which are designed to construct scientifically valid tool workflows and ground conclusions in geospatial evidence. The challenge for these agents lies in navigating complex constraints related to sensing semantics, product dependencies, and spatial/temporal compatibility. Existing systems often struggle with broad operation spaces or fail to effectively organize heterogeneous knowledge for reuse. GeoForge addresses these issues by operating as a training-free, self-evolving system. It converts completed task trajectories into a structured, non-parametric execution state. The framework utilizes three complementary memory types—Workflow Graph Memory, Action-Level Experiences, and Adapted Skill Standard Operating Procedure—to retrieve task-conditioned priors that guide tool execution. After each task, a safety-gated distillation process converts grounded trajectories into reusable knowledge, allowing the system to improve its planning and reasoning without requiring updates to the underlying Large Language Model. Experimental results across various geospatial benchmarks demonstrate that GeoForge consistently boosts task accuracy and tool-use trajectory quality, while substantially reducing planning and reasoning errors for most LLM backbones. This represents a significant step towards more robust and adaptable AI agents for complex scientific domains.