Project Lethe: A Bio-Inspired Cognitive Filter for Sub-Microsecond Continual Learning and Active Forgetting in Edge AI

This repository contains the research paper for Project Lethe, a novel bio-inspired cognitive architecture designed to solve the two fundamental limitations of modern Artificial Intelligence at the edge: catastrophic forgetting and compute inefficiency. Abstract: Modern Artificial Intelligence systems, particularly deep neural networks, suffer from catastrophic forgetting during continual learning and massive computational inefficiency when processing highly predictable data. This paper introduces Project Lethe, a bio-inspired cognitive architecture designed to mimic the human brain's dopamine-driven sensory gating and synaptic decay. Lethe operates entirely in O(1) memory per dimension and O(D) compute per vector, allowing it to act as an ultra-fast "sensory firewall" for heavier AI pipelines. By utilizing a dynamically habituating Diagonal Mahalanobis approximation for surprisal estimation, Lethe drops up to 90% of routine noise in nanoseconds. Surprising anomalies are buffered and consolidated offline into Long-Term Memory (LTM) via weighted concept generalization, preventing catastrophic forgetting. Finally, passive synaptic decay ensures unbounded memory growth is mitigated through true biological forgetting. Dataset: Evaluated on the NASA Numenta Anomaly Benchmark (NAB).Extreme Performance: Achieved a throughput of 1.89 Million vectors per second with sub-microsecond latency (~500ns) natively compiled in Rust.Sensory Gating (Compression): Successfully dropped 56% of routine background noise dynamically.Semantic Consolidation: Condensed nearly 10,000 raw time-series anomalies into exactly 23 Long-Term Semantic Concepts.Biological Forgetting: Implemented a mathematically bounded garbage collector using exponential synaptic decay to prevent infinite memory leaks.