feat(memory): hybrid vector search via embedder

- repository: optional pgvector embedding column (migration with
  CREATE EXTENSION guarded by main), SearchByVector cosine search,
  vector-enabled insert path
- service: SetEmbedder — memories are embedded on Add; Search merges
  vector hits with n-gram keyword matches (vector first, deduped)
- embedder was already implemented (OpenAI-compatible API + n-gram FNV
  fallback) but never instantiated; main now wires it from
  EMBEDDING_API_URL/EMBEDDING_MODEL config

backend/services/memory/service.go

@@ -10,7 +10,8 @@ import (
 
 // Service handles business logic for the memory system.
 type Service struct {
-	repo *Repository
+	repo     *Repository
+	embedder *Embedder
 }
 
 // NewService creates a new memory service.
@@ -18,6 +19,9 @@ func NewService(repo *Repository) *Service {
 	return &Service{repo: repo}
 }
 
+// SetEmbedder enables embedding-based semantic search.
+func (s *Service) SetEmbedder(e *Embedder) { s.embedder = e }
+
 // Add stores a new memory with deduplication based on keyword overlap.
 func (s *Service) Add(ctx context.Context, userID, sourceID, content, category string, priority float64) (*Memory, error) {
 	if priority <= 0 { priority = 0.5 }
@@ -58,6 +62,11 @@ func (s *Service) Add(ctx context.Context, userID, sourceID, content, category s
 		Priority: priority,
 		Decay:    1.0,
 	}
+	if s.embedder != nil {
+		if vec, err := s.embedder.Embed(ctx, content); err == nil {
+			m.Embedding = vec
+		}
+	}
 
 	if _, err := s.repo.Create(ctx, m); err != nil {
 		return nil, fmt.Errorf("add memory: %w", err)
@@ -65,23 +74,46 @@ func (s *Service) Add(ctx context.Context, userID, sourceID, content, category s
 	return m, nil
 }
 
-// Search performs combined keyword search.
+// Search performs hybrid search: vector similarity (when an embedder is
+// configured) merged with n-gram keyword matches, vector hits first.
 func (s *Service) Search(ctx context.Context, q SearchQuery) ([]Memory, error) {
 	if q.Limit <= 0 { q.Limit = 10 }
 
-	// Use n-gram based search for better Chinese text matching
-	nGrams := extractNGrams(q.Query, 2)
+	seen := make(map[string]bool)
 	var memories []Memory
-	var err error
 
+	if s.embedder != nil {
+		if vec, err := s.embedder.Embed(ctx, q.Query); err == nil {
+			if vecHits, err := s.repo.SearchByVector(ctx, q.UserID, vec, q.Limit); err == nil {
+				for _, m := range vecHits {
+					memories = append(memories, m)
+					seen[m.ID] = true
+				}
+			}
+		}
+	}
+
+	// N-gram keyword search for better Chinese text matching
+	nGrams := extractNGrams(q.Query, 2)
+	var kwHits []Memory
+	var err error
 	if len(nGrams) > 0 {
-		memories, err = s.repo.SearchByNGrams(ctx, q.UserID, nGrams[:min(3, len(nGrams))], q.Limit)
+		kwHits, err = s.repo.SearchByNGrams(ctx, q.UserID, nGrams[:min(3, len(nGrams))], q.Limit)
 	} else {
-		memories, err = s.repo.ListByUser(ctx, q.UserID, q.Limit)
+		kwHits, err = s.repo.ListByUser(ctx, q.UserID, q.Limit)
 	}
 	if err != nil {
 		return nil, fmt.Errorf("search memory: %w", err)
 	}
+	for _, m := range kwHits {
+		if !seen[m.ID] {
+			memories = append(memories, m)
+			seen[m.ID] = true
+		}
+	}
+	if len(memories) > q.Limit {
+		memories = memories[:q.Limit]
+	}
 
 	for _, m := range memories {
 		go s.repo.RecordAccess(context.Background(), m.ID)