Fix_Semantic_Retrieval_Threshold

config.yaml

@@ -13,7 +13,7 @@ vector:
   # Using a smaller model to avoid OOM on standard machines during demo
   embedding_model: "all-MiniLM-L6-v2"
   dim: 384
-  similarity_threshold: 0.4 # Slightly lower for MiniLM
+  similarity_threshold: 0.5 # Filter out weak semantic matches (was 0.4)
 evaluation:
   checkpoints: [100, 500, 937, 1000, 1200]
   recall_k: 6

src/neurohack_memory/system.py

@@ -108,10 +108,18 @@ class MemorySystem:
 
     def retrieve(self, query):
         cfgm = self.cfg["memory"]
+        cfgv = self.cfg.get("vector", {})
+        min_similarity = cfgv.get("similarity_threshold", 0.4)
+        
         t = Timer.start()
         hits = self.vindex.search(query, top_k=max(10, cfgm["top_k"]*3))
         candidates = []
         for mid, base_score in hits:
+            # CRITICAL FIX: Filter out semantically irrelevant matches
+            # If base_score (cosine similarity) is below threshold, skip it
+            if base_score < min_similarity:
+                continue
+                
             m = self._memory_cache.get(mid)
             if not m:
                 continue