feat: make similarity thresholds configurable for judge and temporal search

src/agents/judge.py

@@ -32,6 +32,7 @@
     OperationType,
 )
 from src.storage.base import BaseVectorStore, SearchResult
+from src.config import settings
 
 
 # ---------------------------------------------------------------------------
@@ -87,13 +88,13 @@ def _format_similar_block(
     return "\n".join(lines)
 
 
-SUMMARY_JUDGE_SIMILARITY_THRESHOLD = 0.4
-
 
 def _has_summary_judge_candidates(
     matches_per_item: Dict[str, List[SearchResult]],
-    threshold: float = SUMMARY_JUDGE_SIMILARITY_THRESHOLD,
+    threshold: Optional[float] = None,
 ) -> bool:
+    if threshold is None:
+        threshold = settings.summary_judge_similarity_threshold
     for matches in matches_per_item.values():
         for match in matches:
             if match.score >= threshold:
@@ -112,11 +113,12 @@ def _filter_matches_by_threshold(
 
 
 def _deterministic_summary_add(items_strings: List[str], confidence: float = 0.8) -> JudgeResult:
+    threshold = settings.summary_judge_similarity_threshold
     operations = [
         Operation(
             type=OperationType.ADD,
             content=item,
-            reason="No similar summary at or above 0.4 — defaulting to ADD.",
+            reason=f"No similar summary at or above {threshold} — defaulting to ADD.",
         )
         for item in items_strings
         if str(item).strip()
@@ -196,7 +198,7 @@ async def arun(self, state: Dict[str, Any]) -> JudgeResult:
         if domain == JudgeDomain.SUMMARY:
             matches_per_item = _filter_matches_by_threshold(
                 matches_per_item,
-                SUMMARY_JUDGE_SIMILARITY_THRESHOLD,
+                settings.summary_judge_similarity_threshold,
             )
 
         # 3. Build the prompt

src/config/settings.py

@@ -136,6 +136,18 @@ class Settings(BaseSettings):
         default=0.4,
         description="LLM temperature for generation"
     )
+    summary_judge_similarity_threshold: float = Field(
+        default=0.4,
+        ge=0.0,
+        le=1.0,
+        description="Threshold score for the Judge to match summary memories"
+    )
+    temporal_search_similarity_threshold: float = Field(
+        default=0.3,
+        ge=-1.0,
+        le=1.0,
+        description="Minimum cosine similarity threshold score for Neo4j temporal search"
+    )
     llm_timeout_seconds: float = Field(
         default=45.0,
         description="Per-agent LLM call timeout in seconds",

src/graph/neo4j_client.py

@@ -30,6 +30,7 @@
 from neo4j import GraphDatabase
 
 from src.graph.schema import GraphSchema
+from src.config import settings
 
 logger = logging.getLogger("xmem.graph.neo4j")
 
@@ -250,7 +251,7 @@ def search_events_by_embedding(
         user_id: str,
         query_text: str,
         top_k: int = 1,
-        similarity_threshold: float = 0.3,
+        similarity_threshold: Optional[float] = None,
     ) -> List[Dict[str, Any]]:
         """Semantic search over event embeddings stored on HAS_EVENT relationships.
 
@@ -263,6 +264,9 @@ def search_events_by_embedding(
         ``similarity_score`` is raw cosine in [-1, 1] (matches the previous
         dot-product semantics, which assumed unit-normalised embeddings).
         """
+        if similarity_threshold is None:
+            similarity_threshold = settings.temporal_search_similarity_threshold
+
         if not self._embedding_fn:
             logger.warning("No embedding function — cannot search by embedding.")
             return []