Temporal Activation Maps: A Lightweight Architecture for Experience-Driven Adaptation in Large Language Models

This paper proposes Temporal Activation Maps (TAM), a lightweight metadata layer that tracks parameter cluster activation patterns across inference events in large language models, enabling experience-driven computational prioritization without modifying model weights. TAM introduces a three-tier consolidation hierarchy mirroring biological memory systems and provides a procedural memory complement to existing declarative memory systems. Estimated storage overhead is 80–230 MB per deployment context. The paper presents the architecture, differentiates TAM from MoE, LoRA, RAG, and sparse autoencoders, estimates resource requirements, and proposes experimental validation methods. Co-developed with an AI system (Claude, Anthropic Opus 4.6).