Hierarchical Memory System with Weighting for Lifelong Learning

We propose a Hierarchical Weighted Memory (HWM) architecture for continual (lifelong) learning that organizes experience across multiple abstraction layers. Inspired by cognitive and neuroscientific principles, HWM aligns with the “Epistemological Building Blocks” framework (Existence → Change → Abstraction → Meta-Abstraction) by dedicating separate memory modules to raw instances, temporal dynamics, conceptual categories, and higher-order rules. Each memory entry carries an importance weight that guides consolidation, replay, and retrieval, thereby prioritizing salient knowledge and mitigating catastrophic forgetting. Our design integrates mechanisms from complementary learning systems and memory-augmented networks: for example, fast episodic memory feeds slower semantic memory (as in Parisi et al., 2018), and parameter-importance weighting (as in EWC and Synaptic Intelligence) governs plasticity. We describe concrete implementation options (e.g. dynamic key-value stores, recurrent transformers) and propose experiments on continual-classification and sequential-reasoning benchmarks. We expect HWM to enable richer knowledge retention and abstraction (beyond what single-layer replay or regularization can achieve), while discussing trade-offs and alternatives such as dual-memory or rehearsal approaches