feat(recall): add metadata sidecar search

app.py

@@ -179,6 +179,7 @@ def _parse_viewer_allowed_origins() -> Any:
 from automem.search.runtime_keywords import load_keyword_runtime
 from automem.search.runtime_recall_helpers import (
     _graph_keyword_search,
+    _metadata_keyword_search,
     _result_passes_filters,
     _vector_filter_only_tag_search,
     _vector_search,

automem/api/recall.py

@@ -15,6 +15,7 @@
     FILTERABLE_RELATIONS,
     RECALL_ADAPTIVE_FLOOR,
     RECALL_EXPANSION_LIMIT,
+    RECALL_METADATA_SEARCH_ENABLED,
     RECALL_MIN_SCORE,
     RECALL_RELATION_LIMIT,
     canonicalize_relation_type,
@@ -1417,6 +1418,7 @@ def handle_recall(
     expansion_limit_default: Optional[int] = None,
     on_access: Optional[Callable[[List[str]], None]] = None,
     jit_enrich_fn: Optional[Callable[[str, Dict[str, Any]], Optional[Dict[str, Any]]]] = None,
+    metadata_keyword_search: Optional[Callable[..., List[Dict[str, Any]]]] = None,
 ):
     query_start = time.perf_counter()
     query_text = (request.args.get("query") or "").strip()
@@ -1666,6 +1668,27 @@ def _run_single_query(
             )
             local_results.extend(graph_matches[:remaining_slots])
 
+        if (
+            graph is not None
+            and metadata_keyword_search is not None
+            and query_str
+            and RECALL_METADATA_SEARCH_ENABLED
+        ):
+            metadata_slots = max(1, min(per_query_limit, 10))
+            metadata_matches = metadata_keyword_search(
+                graph,
+                query_str,
+                metadata_slots,
+                local_seen,
+                start_time=start_time,
+                end_time=end_time,
+                tag_filters=tag_filters,
+                tag_mode=tag_mode,
+                tag_match=tag_match,
+                exclude_tags=exclude_tags,
+            )
+            local_results.extend(metadata_matches)
+
         tags_only_request = (
             not query_str
             and not (embedding_param and embedding_param.strip())
@@ -2139,6 +2162,7 @@ def create_recall_blueprint(
     summarize_relation_node: Callable[[Dict[str, Any]], Dict[str, Any]] | None = None,
     on_access: Optional[Callable[[List[str]], None]] = None,
     jit_enrich_fn: Optional[Callable[[str, Dict[str, Any]], Optional[Dict[str, Any]]]] = None,
+    metadata_keyword_search: Optional[Callable[..., List[Dict[str, Any]]]] = None,
 ) -> Blueprint:
     bp = Blueprint("recall", __name__)
 
@@ -2170,6 +2194,7 @@ def recall_memories() -> Any:
             expansion_limit_default=RECALL_EXPANSION_LIMIT,
             on_access=on_access,
             jit_enrich_fn=jit_enrich_fn,
+            metadata_keyword_search=metadata_keyword_search,
         )
 
     @bp.route("/startup-recall", methods=["GET"])

automem/api/runtime_bootstrap.py

@@ -68,6 +68,7 @@ def register_blueprints(
     consolidation_tick_seconds: int,
     consolidation_history_limit: int,
     require_api_token_fn: Callable[[], None],
+    metadata_keyword_search_fn: Optional[Callable[..., list[dict[str, Any]]]] = None,
 ) -> None:
     health_bp = create_health_blueprint(
         get_memory_graph_fn,
@@ -106,6 +107,7 @@ def register_blueprints(
         summarize_relation_node_fn,
         update_last_accessed_fn,
         jit_enrich_fn=jit_enrich_fn,
+        metadata_keyword_search=metadata_keyword_search_fn,
     )
 
     memory_bp = create_memory_blueprint_full(

automem/api/runtime_recall_routes.py

@@ -28,6 +28,7 @@ def recall_memories(
     emit_event_fn: Any,
     utc_now_fn: Any,
     abort_fn: Any,  # noqa: ARG001 - kept for DI compatibility
+    metadata_keyword_search_fn: Any = None,
 ) -> Any:
     query_start = perf_counter_fn()
     query_text = (request_obj.args.get("query") or "").strip()
@@ -66,6 +67,7 @@ def recall_memories(
         default_expand_relations=default_expand_relations,
         relation_limit=recall_relation_limit,
         expansion_limit_default=recall_expansion_limit,
+        metadata_keyword_search=metadata_keyword_search_fn,
     )
 
     elapsed_ms = int((perf_counter_fn() - query_start) * 1000)

automem/config.py

@@ -139,6 +139,11 @@
 RECALL_EXPANSION_LIMIT = int(os.getenv("RECALL_EXPANSION_LIMIT", "25"))
 RECALL_MIN_SCORE = float(os.getenv("RECALL_MIN_SCORE", "0.0"))
 RECALL_ADAPTIVE_FLOOR = os.getenv("RECALL_ADAPTIVE_FLOOR", "true").lower() in ("true", "1", "yes")
+RECALL_METADATA_SEARCH_ENABLED = os.getenv("RECALL_METADATA_SEARCH_ENABLED", "true").lower() not in {
+    "0",
+    "false",
+    "no",
+}
 
 # Memory content size limits (governs auto-summarization on store)
 # Soft limit: Content above this triggers auto-summarization
@@ -447,6 +452,7 @@ def expand_relation_query_types(relation_types: Iterable[str]) -> list[str]:
 # Search weighting parameters (can be overridden via environment variables)
 SEARCH_WEIGHT_VECTOR = float(os.getenv("SEARCH_WEIGHT_VECTOR", "0.35"))
 SEARCH_WEIGHT_KEYWORD = float(os.getenv("SEARCH_WEIGHT_KEYWORD", "0.35"))
+SEARCH_WEIGHT_METADATA = float(os.getenv("SEARCH_WEIGHT_METADATA", "0.35"))
 SEARCH_WEIGHT_TAG = float(os.getenv("SEARCH_WEIGHT_TAG", "0.2"))
 SEARCH_WEIGHT_IMPORTANCE = float(os.getenv("SEARCH_WEIGHT_IMPORTANCE", "0.1"))
 SEARCH_WEIGHT_CONFIDENCE = float(os.getenv("SEARCH_WEIGHT_CONFIDENCE", "0.05"))

automem/runtime_wiring.py

@@ -77,6 +77,7 @@ def wire_recall_and_blueprints(
         consolidation_tick_seconds=module.CONSOLIDATION_TICK_SECONDS,
         consolidation_history_limit=module.CONSOLIDATION_HISTORY_LIMIT,
         require_api_token_fn=module.require_api_token,
+        metadata_keyword_search_fn=module._metadata_keyword_search,
     )