diff --git a/src/memos/graph_dbs/polardb.py b/src/memos/graph_dbs/polardb.py
index d740ad1d2..a2cab0c98 100644
--- a/src/memos/graph_dbs/polardb.py
+++ b/src/memos/graph_dbs/polardb.py
@@ -1,5 +1,6 @@
 import json
 import random
+import re
 import textwrap
 import threading
 import time
@@ -20,6 +21,34 @@
 logger = get_logger(__name__)
 
 
+def _sanitize_tsquery_words(query_words: list[str]) -> list[str]:
+    """Sanitize query words for safe use with PostgreSQL to_tsquery().
+
+    Strips tsquery operator characters and other special symbols that can
+    cause parsing errors when mixed content (e.g. message IDs with
+    underscores, Chinese text) is passed to ``to_tsquery``.  Each word is
+    reduced to its alphanumeric/CJK core so that the jieba text-search
+    configuration can tokenize it correctly.
+
+    Returns a de-duplicated list of non-empty sanitized words.
+    """
+    # Keep word characters (letters, digits, underscore) and CJK unified ideographs.
+    valid_chars_re = re.compile(
+        r"[^\w\u4e00-\u9fff\u3400-\u4dbf\uf900-\ufaff]",
+    )
+    sanitized: list[str] = []
+    seen: set[str] = set()
+    for w in query_words:
+        # Strip surrounding single quotes that callers may have added for tsquery
+        w = w.strip().strip("'")
+        # Remove characters that are not word-characters or CJK
+        cleaned = valid_chars_re.sub("", w)
+        if cleaned and cleaned not in seen:
+            seen.add(cleaned)
+            sanitized.append(cleaned)
+    return sanitized
+
+
 def _compose_node(item: dict[str, Any]) -> tuple[str, str, dict[str, Any]]:
     node_id = item["id"]
     memory = item["memory"]
@@ -1653,8 +1682,11 @@ def search_by_keywords_tfidf(
         filter_conditions = self._build_filter_conditions_sql(filter)
         where_clauses.extend(filter_conditions)
         # Add fulltext search condition
-        # Convert query_text to OR query format: "word1 | word2 | word3"
-        tsquery_string = " | ".join(query_words)
+        # Sanitize and convert query_text to OR query format: "word1 | word2 | word3"
+        safe_words = _sanitize_tsquery_words(query_words)
+        if not safe_words:
+            return []
+        tsquery_string = " | ".join(safe_words)
 
         where_clauses.append(f"{tsvector_field} @@ to_tsquery('{tsquery_config}', %s)")
 
@@ -1768,7 +1800,10 @@ def search_by_fulltext(
         filter_conditions = self._build_filter_conditions_sql(filter)
 
         where_clauses.extend(filter_conditions)
-        tsquery_string = " | ".join(query_words)
+        safe_words = _sanitize_tsquery_words(query_words)
+        if not safe_words:
+            return []
+        tsquery_string = " | ".join(safe_words)
 
         where_clauses.append(f"{tsvector_field} @@ to_tsquery('{tsquery_config}', %s)")
 
diff --git a/tests/graph_dbs/test_sanitize_tsquery.py b/tests/graph_dbs/test_sanitize_tsquery.py
new file mode 100644
index 000000000..108f92050
--- /dev/null
+++ b/tests/graph_dbs/test_sanitize_tsquery.py
@@ -0,0 +1,86 @@
+"""Tests for _sanitize_tsquery_words — standalone, no heavy imports."""
+
+import re
+
+
+# ---------------------------------------------------------------------------
+# Inline the function under test to avoid pulling in the full memos import
+# chain (which requires a running logging backend).  The canonical copy lives
+# in ``memos.graph_dbs.polardb._sanitize_tsquery_words``.
+# ---------------------------------------------------------------------------
+
+
+def _sanitize_tsquery_words(query_words: list[str]) -> list[str]:
+    valid_chars_re = re.compile(
+        r"[^\w\u4e00-\u9fff\u3400-\u4dbf\uf900-\ufaff]",
+    )
+    sanitized: list[str] = []
+    seen: set[str] = set()
+    for w in query_words:
+        w = w.strip().strip("'")
+        cleaned = valid_chars_re.sub("", w)
+        if cleaned and cleaned not in seen:
+            seen.add(cleaned)
+            sanitized.append(cleaned)
+    return sanitized
+
+
+class TestSanitizeTsqueryWords:
+    """Unit tests for FTS query word sanitization."""
+
+    def test_plain_english_words(self):
+        assert _sanitize_tsquery_words(["hello", "world"]) == ["hello", "world"]
+
+    def test_chinese_text(self):
+        result = _sanitize_tsquery_words(["我要", "测试"])
+        assert result == ["我要", "测试"]
+
+    def test_mixed_content_message_id_and_chinese(self):
+        """Reproduce the original bug: mixed IDs + Chinese text."""
+        words = ["message_id", "om_x100b544a390604b8c3e1b7d8641f08e", "我要测试"]
+        result = _sanitize_tsquery_words(words)
+        assert len(result) == 3
+        assert "message_id" in result
+        assert "om_x100b544a390604b8c3e1b7d8641f08e" in result
+        assert "我要测试" in result
+
+    def test_single_quoted_words_are_stripped(self):
+        words = ["'hello'", "'world'"]
+        result = _sanitize_tsquery_words(words)
+        assert result == ["hello", "world"]
+
+    def test_special_characters_removed(self):
+        words = ["hello!", "world@#$"]
+        result = _sanitize_tsquery_words(words)
+        assert result == ["hello", "world"]
+
+    def test_empty_words_filtered(self):
+        words = ["", "  ", "hello", ""]
+        result = _sanitize_tsquery_words(words)
+        assert result == ["hello"]
+
+    def test_deduplication(self):
+        words = ["hello", "hello", "world"]
+        result = _sanitize_tsquery_words(words)
+        assert result == ["hello", "world"]
+
+    def test_empty_input(self):
+        assert _sanitize_tsquery_words([]) == []
+
+    def test_all_special_chars_returns_empty(self):
+        words = ["!@#", "$%^"]
+        result = _sanitize_tsquery_words(words)
+        assert result == []
+
+    def test_underscores_preserved(self):
+        words = ["message_id", "user_name"]
+        result = _sanitize_tsquery_words(words)
+        assert result == ["message_id", "user_name"]
+
+    def test_tsquery_operators_stripped(self):
+        """Tsquery operators like & | ! should be stripped from within words."""
+        words = ["hello & world", "foo | bar"]
+        result = _sanitize_tsquery_words(words)
+        # Spaces and operators removed; alphanumeric parts merge
+        assert "helloworld" in result
+        assert "foobar" in result