Active Dreaming Memory: Biologically-Inspired Episodic Consolidation for Lifelong Learning in Autonomous Agents

Abstract

Autonomous agents powered by Large Language Models (LLMs) often suffer from catastrophic forgetting and lack temporal continuity. We present Active Dreaming Memory (ADM), a biologically-inspired dual-store memory system that enables agents to learn from execution failures without fine-tuning. Our key innovation is a counterfactual verification mechanism that validates candidate rules through synthetic scenario simulation before committing them to long-term memory. Experiments across six diverse domains demonstrate that ADM improves first-episode learning efficiency by 2× and achieves 83% success rate, significantly outperforming Reflexion, MemGPT, and Self-RAG. Catastrophic forgetting stress tests show 95% retention after 500 episodes. Theoretical analysis proves bounded forgetting guarantees and logarithmic memory growth. We release open-source code to support reproducible lifelong learning research.