diff --git a/pyrit/memory/azure_sql_memory.py b/pyrit/memory/azure_sql_memory.py
index 080fe6bde8..10775019d2 100644
--- a/pyrit/memory/azure_sql_memory.py
+++ b/pyrit/memory/azure_sql_memory.py
@@ -737,7 +737,10 @@ def _execute_get_conversation_stats(self, *, conversation_ids: Sequence[str]) ->
                             p2.converted_value_data_type AS last_data_type
                         FROM "PromptMemoryEntries" p2
                         WHERE p2.conversation_id = aggregate_rows.conversation_id
-                        ORDER BY p2.sequence DESC, p2.id DESC
+                        ORDER BY
+                            p2.sequence DESC,
+                            p2.timestamp DESC,
+                            LOWER(CAST(p2.id AS VARCHAR(36))) COLLATE Latin1_General_100_BIN2 DESC
                     ) AS latest
                     """
                 ).columns(created_at=UTCDateTime())
diff --git a/pyrit/memory/sqlite_memory.py b/pyrit/memory/sqlite_memory.py
index 1756c579c4..ef5cf2e114 100644
--- a/pyrit/memory/sqlite_memory.py
+++ b/pyrit/memory/sqlite_memory.py
@@ -800,7 +800,7 @@ def _execute_get_conversation_stats(self, *, conversation_ids: Sequence[str]) ->
                             SELECT p2.id
                             FROM "PromptMemoryEntries" p2
                             WHERE p2.conversation_id = aggregate_rows.conversation_id
-                            ORDER BY p2.sequence DESC, p2.id DESC
+                            ORDER BY p2.sequence DESC, p2.timestamp DESC, p2.id DESC
                             LIMIT 1
                         )
                     """
diff --git a/tests/unit/memory/test_azure_sql_memory.py b/tests/unit/memory/test_azure_sql_memory.py
index a2e5b8a8bf..dc445517fc 100644
--- a/tests/unit/memory/test_azure_sql_memory.py
+++ b/tests/unit/memory/test_azure_sql_memory.py
@@ -621,6 +621,27 @@ def test_get_conversation_stats_types_created_at_as_utc_datetime(
     assert isinstance(sql.selected_columns.created_at.type, UTCDateTime)
 
 
+def test_get_conversation_stats_orders_the_last_piece_canonically(
+    uninitialized_memory_interface: AzureSQLMemory,
+) -> None:
+    """The last message lookup orders pieces the way the conversation page does.
+
+    SQL Server UUID ordering differs from the SDK string comparison the page
+    uses, so the lookup sorts ids through the same lowered collation, and the
+    timestamp decides between pieces that share a sequence number.
+    """
+    session = MagicMock()
+    session.execute.return_value.fetchall.return_value = []
+
+    with patch.object(uninitialized_memory_interface, "_get_session", return_value=session):
+        uninitialized_memory_interface._execute_get_conversation_stats(conversation_ids=["conversation"])
+
+    sql = str(session.execute.call_args.args[0])
+    assert "p2.sequence DESC" in sql
+    assert "p2.timestamp DESC" in sql
+    assert "LOWER(CAST(p2.id AS VARCHAR(36))) COLLATE Latin1_General_100_BIN2 DESC" in sql
+
+
 def test_get_conversation_stats_uses_one_latest_row_apply(
     uninitialized_memory_interface: AzureSQLMemory,
 ) -> None:
diff --git a/tests/unit/memory/test_sqlite_memory.py b/tests/unit/memory/test_sqlite_memory.py
index 1590718dde..1e7ab50719 100644
--- a/tests/unit/memory/test_sqlite_memory.py
+++ b/tests/unit/memory/test_sqlite_memory.py
@@ -845,7 +845,7 @@ def capture_statement(conn, cursor, statement, parameters, context, executemany)
 
     sql = "\n".join(statements).upper()
     assert 'LEFT JOIN "PROMPTMEMORYENTRIES" LATEST' in sql
-    assert "ORDER BY P2.SEQUENCE DESC, P2.ID DESC" in sql
+    assert "ORDER BY P2.SEQUENCE DESC, P2.TIMESTAMP DESC, P2.ID DESC" in sql
     assert "ROW_NUMBER" not in sql
 
 
@@ -928,6 +928,52 @@ async def test_get_conversation_stats_does_not_read_labels_from_prompt_entries(s
     assert result[conv_id].labels == {}
 
 
+async def test_get_conversation_stats_picks_the_canonical_last_piece(sqlite_instance):
+    """The preview names the piece the conversation page orders last.
+
+    Pieces of one turn share their sequence number, and the page orders them
+    by timestamp before the id tiebreak, so the preview lookup has to honor
+    the same order instead of handing the win to the bigger id.
+    """
+    import uuid
+    from datetime import UTC, datetime, timedelta
+
+    from pyrit.models import MessagePiece
+
+    conv_id = str(uuid.uuid4())
+    base = datetime(2026, 1, 1, tzinfo=UTC)
+    earlier = MessagePiece(
+        id=uuid.UUID("ffffffff-0000-0000-0000-000000000001"),
+        role="assistant",
+        original_value="earlier piece",
+        original_value_data_type="text",
+        converted_value="earlier piece",
+        converted_value_data_type="text",
+        conversation_id=conv_id,
+        sequence=2,
+        timestamp=base,
+    )
+    later = MessagePiece(
+        id=uuid.UUID("00000000-0000-0000-0000-000000000002"),
+        role="assistant",
+        original_value="later piece",
+        original_value_data_type="text",
+        converted_value="later piece",
+        converted_value_data_type="text",
+        conversation_id=conv_id,
+        sequence=2,
+        timestamp=base + timedelta(seconds=10),
+    )
+    for piece in (earlier, later):
+        sqlite_instance._insert_entry(PromptMemoryEntry(entry=piece))
+
+    stats = await sqlite_instance.get_conversation_stats_async(conversation_ids=[conv_id])
+    page = await sqlite_instance.get_conversation_message_page_async(conversation_id=conv_id, limit=100)
+
+    assert stats[conv_id].last_message_preview == "later piece"
+    assert page.messages[-1].message_pieces[-1].converted_value == "later piece"
+
+
 async def test_get_conversation_stats_preview_caps_raw_value_at_fetch_limit(sqlite_instance):
     """Memory caps the raw last_message_preview at PREVIEW_FETCH_MAX_LEN.