A new memory system for autonomous agents addresses the "curse of dimensionality" by applying Independent Component Analysis (ICA) to reduce data into five components, then compressing dimensions based on short-term memory variance to encourage behavioral variety.
Episodes are stored in a FAISS vector database using SONAR encoder embeddings and undergo consolidation and generalization at each run. ICA identifies normally distributed, independent components, with tests showing that reducing data to three dimensions yields interpretable fields for activity, hierarchy, and social states. The system calculates component differences in short-term memory; low variance triggers compression of the corresponding vector database dimension, making Manhattan distance retrieval more likely to yield episodes spanning the full spectrum of that component.
Testing demonstrates that this mechanism alters agent behavior dynamically, such as shifting from fieldwork to smelting or prioritizing task completion when anxiety components rise, driven by compressed dimensions rather than explicit initialization.