Python: extract keywords from non-English text for topic selection

python/packages/core/agent_framework/_harness/_memory.py

@@ -67,7 +67,10 @@
 """
 _FILE_HISTORY_ENCODED_SESSION_PREFIX = "~session-"
 HistoryMessageFilter = Callable[[Message], Message | None]
-_WORD_PATTERN = re.compile(r"[a-z0-9][a-z0-9_-]{1,}", flags=re.IGNORECASE)
+# Match word-like tokens of >=2 characters. ``[^\W_]`` is a Unicode letter or
+# digit (excluding underscore), so CJK/Cyrillic/etc. text yields keywords too
+# instead of nothing -- otherwise non-English messages never match topic files.
+_WORD_PATTERN = re.compile(r"[^\W_][\w-]+")
 
 
 def _payload_preview(text: str, *, limit: int = 120) -> str:

python/packages/core/tests/core/test_harness_memory.py

@@ -861,3 +861,21 @@ async def test_memory_extraction_propagates_programmer_errors(tmp_path) -> None:
             context=context,
             now=datetime(2026, 4, 22, tzinfo=timezone.utc),
         )
+
+
+def test_extract_keywords_handles_non_english_text() -> None:
+    """Keyword extraction yields tokens for non-English input (#6989).
+
+    The pattern used to be ASCII-only, so CJK/Cyrillic messages produced an
+    empty keyword set and therefore never matched any topic file.
+    """
+    from agent_framework._harness._memory import _extract_keywords  # pyright: ignore[reportPrivateUsage]
+
+    cjk = _extract_keywords([Message(role="user", contents=["こんにちは 元気ですか"])])
+    cyrillic = _extract_keywords([Message(role="user", contents=["привет мир друзья"])])
+    english = _extract_keywords([Message(role="user", contents=["hello world"])])
+
+    assert cjk == {"こんにちは", "元気ですか"}
+    assert cyrillic == {"привет", "мир", "друзья"}
+    # English extraction is unchanged.
+    assert english == {"hello", "world"}