From 7d3038510c40e55f09c67b24e4c2cdcf6e904d9c Mon Sep 17 00:00:00 2001
From: simpleqt <89645338+simpleqt@users.noreply.github.com>
Date: Wed, 30 Sep 2026 23:58:52 +0800
Subject: [PATCH] Only split bracketed lists whose items are reference tags

split_continuous_references treated any [a, b] block as a reference
list, corrupting plain prose like 'the set is [apple, banana]'. The
items must now match int:id reference-tag shapes.

Fixes #2446
---
 src/memos/mem_os/utils/reference_utils.py  |  8 ++++++
 tests/mem_os/test_reference_split_shape.py | 29 ++++++++++++++++++++++
 2 files changed, 37 insertions(+)
 create mode 100644 tests/mem_os/test_reference_split_shape.py

diff --git a/src/memos/mem_os/utils/reference_utils.py b/src/memos/mem_os/utils/reference_utils.py
index 09b812207..de2d744e4 100644
--- a/src/memos/mem_os/utils/reference_utils.py
+++ b/src/memos/mem_os/utils/reference_utils.py
@@ -41,6 +41,14 @@ def split_continuous_references(text: str) -> str:
     # Check if there's a comma between brackets
     if "," not in content_between_brackets:
         return text
+    # Every item must look like a reference tag "int:id"; plain bracketed
+    # prose such as "[apple, banana]" must pass through untouched
+    import re
+
+    if not re.fullmatch(
+        r"\s*\d+:[^,\s]+(?:,\s*\d+:[^,\s]+)*\s*", 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(",", "]["))
 
diff --git a/tests/mem_os/test_reference_split_shape.py b/tests/mem_os/test_reference_split_shape.py
new file mode 100644
index 000000000..56edb89eb
--- /dev/null
+++ b/tests/mem_os/test_reference_split_shape.py
@@ -0,0 +1,29 @@
+import pytest
+
+from memos.mem_os.utils.reference_utils import split_continuous_references
+
+
+@pytest.mark.parametrize(
+    ("text", "expected"),
+    [
+        ("[1:92ff35fb, 4:bfe6f044]", "[1:92ff35fb][4:bfe6f044]"),
+        ("[1:92ff35fb,4:bfe6f044]", "[1:92ff35fb][4:bfe6f044]"),
+        ("See [1:aa, 2:bb] now", "See [1:aa][2:bb] now"),
+    ],
+)
+def test_reference_lists_are_split(text, expected):
+    assert split_continuous_references(text) == expected
+
+
+@pytest.mark.parametrize(
+    "text",
+    [
+        "The set is [apple, banana] here",
+        "[1, 2, 3]",
+        "[some note, another note]",
+        "no brackets at all",
+    ],
+)
+def test_non_reference_text_passes_through(text):
+    # plain bracketed prose used to be corrupted into [apple][banana]
+    assert split_continuous_references(text) == text