Built an AI memory system based on cognitive science instead of vector databases

Most AI agent memory is just vector DB + semantic search. everything, retrieve by similarity. It works, but it doesn't scale well over time. The noise floor keeps rising and recall quality degrades. I took a different approach and built memory using actual cognitive science models. ACT-R activation decay, Hebbian learning, Ebbinghaus forgetting curves. The system actively forgets stale information and reinforces frequently-used memories, like how human memory works. After 30 days in production: 3,846 memories, 230K+ recalls, $0 inference cost (pure Python, no embeddings required.