🧠 3D Agent Memory Retrieval: Recency Decay, Cosine Relevance, Importance Weights and Top-k Recall
A Blender-built memory vault with one bay per topic: 240 cards placed by age and importance, a query terminal and a score = recency + relevance + importance rule; move the weights and the half-life and watch top-k precision and recall change, checked against heap selection, Monte Carlo and closed forms.
⚙ Under the hood
A Blender-built memory vault with one bay per topic: 240 cards placed by age and importance, a query terminal and a score = recency + relevance + importance rule; move the weights and the half-life and watch top-k precision and recall change, checked against heap selection, Monte Carlo and closed forms.
Agent MemoryRetrievalCosine SimilarityRecency DecayImportance ScoreTop-kPrecision and RecallEmbeddings
3D · Three.js / WebGL renderer · 60 FPS target · runs fully client-side, no install