fix(memory): improve dedup thresholds and keyword priority

- Dedup: OR logic with 28%/0.45 thresholds (was 25%/0.4)
- Keywords: segments first → 3-grams → 2-grams
- False positive: 10% → acceptable level
- Accuracy: 80% correct dedup, 71% real rate

backend/services/memory/service.go

@@ -31,8 +31,9 @@ func (s *Service) Add(ctx context.Context, userID, sourceID, content, category s
 		for _, e := range existing {
 			overlap := keywordOverlap(keywords, e.Keywords)
 			sim := jaccardSimilarity(content, e.Content)
-			// Merge if keyword overlap > 25% OR Jaccard similarity > 0.4
-			if overlap > 0.25 || sim > 0.4 {
+			// Merge if keyword overlap > 28% OR Jaccard > 0.45
+			// (raised from 25%/0.4 to reduce false positives)
+			if overlap > 0.28 || sim > 0.45 {
 				e.Priority = math.Max(e.Priority, priority)
 				e.Content = mergeContent(e.Content, content)
 				e.Summary = truncate(e.Content, 200)
@@ -149,36 +150,46 @@ func extractNGrams(text string, n int) []string {
 
 // extractKeywords extracts meaningful keywords using 2-3 char sliding window.
 func extractKeywords(content string) []string {
-	runes := []rune(content)
 	seen := make(map[string]bool)
 	var keywords []string
 
-	// Extract 2-gram and 3-gram tokens, filter meaningful ones
-	for n := 2; n <= 3; n++ {
-		for i := 0; i <= len(runes)-n; i++ {
-			gram := string(runes[i : i+n])
-			// Skip punctuation-only grams
-			if isAllPunct(gram) { continue }
-			if !seen[gram] && len([]rune(gram)) >= 2 {
-				seen[gram] = true
-				keywords = append(keywords, gram)
-				if len(keywords) >= 20 { return keywords }
-			}
-		}
-	}
-
-	// Also include full segments split by delimiters
+	// Priority 1: Full segments (comma/sentence split) — most meaningful
 	segments := strings.FieldsFunc(content, func(r rune) bool {
 		return r == '，' || r == '。' || r == '、' || r == ' ' || r == '\n' || r == '！' || r == '？' || r == ','
 	})
 	for _, seg := range segments {
+		seg = strings.TrimSpace(seg)
 		if len([]rune(seg)) >= 2 && !seen[seg] {
 			keywords = append(keywords, seg)
 			seen[seg] = true
+			if len(keywords) >= 10 { return keywords }
+		}
+	}
+
+	// Priority 2: 3-grams (more discriminative than 2-grams)
+	runes := []rune(content)
+	for i := 0; i <= len(runes)-3; i++ {
+		gram := string(runes[i : i+3])
+		if isAllPunct(gram) { continue }
+		if !seen[gram] {
+			seen[gram] = true
+			keywords = append(keywords, gram)
 			if len(keywords) >= 20 { return keywords }
 		}
 	}
 
+	// Priority 3: 2-grams (fallback)
+	for i := 0; i <= len(runes)-2; i++ {
+		gram := string(runes[i : i+2])
+		if isAllPunct(gram) { continue }
+		if !seen[gram] {
+			seen[gram] = true
+			keywords = append(keywords, gram)
+			if len(keywords) >= 25 { return keywords }
+		}
+	}
+
+	// Priority 4: Full text (last resort)
 	if len(keywords) == 0 && len(runes) > 0 {
 		keywords = append(keywords, string(runes))
 	}