Evo Memory

This is the persistent learning layer for research workflows. It maintains two memory stores: one for research directions (what's promising vs. what's been tried and failed), another for technical strategies (data processing, training tricks, architecture patterns that actually worked). Three evolution mechanisms update these stores after ideation cycles and experiment runs, classifying failures as implementation bugs versus fundamental dead ends. The paper implementation uses embedding-based retrieval to pull relevant knowledge into new cycles. Use it when you finish an ideation tournament, complete an experiment pipeline, or need to check what you learned last time before starting fresh. It's meta-learning infrastructure, not experiment execution.