diff --git a/src/memos/mem_os/utils/reference_utils.py b/src/memos/mem_os/utils/reference_utils.py
index 09b812207..a843cf024 100644
--- a/src/memos/mem_os/utils/reference_utils.py
+++ b/src/memos/mem_os/utils/reference_utils.py
@@ -1,3 +1,5 @@
+import re
+
 from memos.memories.textual.item import (
     TextualMemoryItem,
 )
@@ -41,8 +43,19 @@ def split_continuous_references(text: str) -> str:
     # Check if there's a comma between brackets
     if "," not in content_between_brackets:
         return text
-    text = text.replace(content_between_brackets, content_between_brackets.replace(", ", "]["))
-    text = text.replace(content_between_brackets, content_between_brackets.replace(",", "]["))
+    # Only reference tags (a numeric id before a colon in every element) are
+    # split; ordinary bracketed text such as "[x, y]" must pass through
+    # untouched.
+    if not re.fullmatch(r"\s*\d+:[^,\s]+(?:,\s*\d+:[^,\s]+)*\s*", content_between_brackets):
+        return text
+    # Split on every comma regardless of the whitespace that follows it: LLM
+    # output mixes "a, b" and "a,b" freely, and the previous two-pass
+    # str.replace handled only one style per call, leaving earlier references
+    # merged when both styles appeared in the same tag.
+    text = text.replace(
+        content_between_brackets,
+        re.sub(r",\s*", "][", content_between_brackets),
+    )
 
     return text
 
@@ -57,8 +70,6 @@ def process_streaming_references_complete(text_buffer: str) -> tuple[str, str]:
     Returns:
         tuple[str, str]: (processed_text, remaining_buffer)
     """
-    import re
-
     # Pattern to match complete reference tags: [refid:memoriesID]
     complete_pattern = r"\[\d+:[^\]]+\]"
 
diff --git a/tests/mem_os/test_reference_utils.py b/tests/mem_os/test_reference_utils.py
new file mode 100644
index 000000000..c1c400b6a
--- /dev/null
+++ b/tests/mem_os/test_reference_utils.py
@@ -0,0 +1,58 @@
+import unittest
+
+from memos.mem_os.utils.reference_utils import split_continuous_references
+
+
+class TestSplitContinuousReferences(unittest.TestCase):
+    def test_spaced_comma(self):
+        self.assertEqual(
+            split_continuous_references("[1:92ff35fb, 4:bfe6f044]"),
+            "[1:92ff35fb][4:bfe6f044]",
+        )
+
+    def test_tight_comma(self):
+        self.assertEqual(
+            split_continuous_references("[1:92ff35fb,4:bfe6f044]"),
+            "[1:92ff35fb][4:bfe6f044]",
+        )
+
+    def test_mixed_separators(self):
+        """Mixed ', ' and ',' separators used to leave earlier refs merged.
+
+        The old two-pass str.replace applied one separator style per call;
+        the first successful pass removed the substring the second pass
+        searched for, so with "[a,b, c]" only the last boundary was split.
+        """
+        self.assertEqual(
+            split_continuous_references("[1:92ff35fb,4:bfe6f044, 7:abcd1234]"),
+            "[1:92ff35fb][4:bfe6f044][7:abcd1234]",
+        )
+
+    def test_ordinary_bracketed_text_untouched(self):
+        # "[x, y]" has no reference-tag shape (numeric id + colon); it must
+        # not be rewritten into "[x][y]" on the streaming chat path
+        self.assertEqual(
+            split_continuous_references("interval [x, y] ends"),
+            "interval [x, y] ends",
+        )
+        self.assertEqual(split_continuous_references("cite [1, 2] here"), "cite [1, 2] here")
+
+    def test_surrounding_text_preserved(self):
+        self.assertEqual(
+            split_continuous_references("see refs [1:aaaa, 2:bbbb] for details"),
+            "see refs [1:aaaa][2:bbbb] for details",
+        )
+
+    def test_non_reference_text_untouched(self):
+        for text in (
+            "",
+            "plain text",
+            "[no commas here]",
+            "two [brackets] twice [here]",
+            "backwards ]here[",
+        ):
+            self.assertEqual(split_continuous_references(text), text)
+
+
+if __name__ == "__main__":
+    unittest.main()