TERRAIN MEMORY: LEVERAGING THE FULL BREADTH OF LLM TRAINING FOR EXPERIENTIAL AI AGENT MEMORY

AI coding agents built on large language models (LLMs) currently access a narrow fraction of their training data—the technical vocabulary of programming—while ignoring the vast majority: literature, nature writing, navigation, warfare, poetry, and the full breadth of human sensory and emotional experience. Simultaneously, all existing agent memory systems operate on a single factual axis, recording what happened without capturing how it felt. We observe that this creates a double waste: the model’s deep understanding of terrain, weather, emotion, and narrative sits dormant, while memory retrieval reduces rich experience to keyword matching. We propose Terrain Memory, an experiential memory architecture for AI agents that encodes task experience as living, multi-dimensional memories anchored to abstract landscape archetypes. Each memory carries terrain (the structural character of the problem), weather (current conditions), sensations (weighted qualitative signals), fears (characteristic risks), and satisfactions (markers of success). Memories follow a living lifecycle: they strengthen through successful application, fade when not revisited, and accumulate into aquifer-level confidence regions that emerge from experience rather than being declared.