diff --git a/CHANGELOG.md b/CHANGELOG.md
index 6bc3a17418b..7d71c1a0d1c 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -14,6 +14,7 @@
 
 ### Bug fixes
 
+* [#7644](https://github.com/rubocop-hq/rubocop/issues/7644): Fix patterns with named wildcards in unions. ([@marcandre][])
 * [#7639](https://github.com/rubocop-hq/rubocop/pull/7639): Fix logical operator edge case in `omit_parentheses` style of `Style/MethodCallWithArgsParentheses`. ([@gsamokovarov][])
 * [#7661](https://github.com/rubocop-hq/rubocop/pull/7661): Fix to return correct info from multi-line regexp. ([@Tietew][])
 * [#7655](https://github.com/rubocop-hq/rubocop/issues/7655): Fix an error when processing a regexp with a line break at the start of capture parenthesis. ([@koic][])
diff --git a/lib/rubocop/node_pattern.rb b/lib/rubocop/node_pattern.rb
index 14e58a05ccb..fd663fbb681 100644
--- a/lib/rubocop/node_pattern.rb
+++ b/lib/rubocop/node_pattern.rb
@@ -188,7 +188,7 @@ def initialize(str, node_var = 'node0')
 
         @temps    = 0  # avoid name clashes between temp variables
         @captures = 0  # number of captures seen
-        @unify    = {} # named wildcard -> temp variable number
+        @unify    = {} # named wildcard -> temp variable
         @params   = 0  # highest % (param) number seen
         run(node_var)
       end
@@ -466,12 +466,67 @@ def insure_same_captures(enum, what)
         end
       end
 
+      def access_unify(name)
+        var = @unify[name]
+
+        if var == :forbidden_unification
+          fail_due_to "Wildcard #{name} was first seen in a subset of a" \
+                      " union and can't be used outside that union"
+        end
+        var
+      end
+
+      def forbid_unification(*names)
+        names.each do |name|
+          @unify[name] = :forbidden_unification
+        end
+      end
+
+      # rubocop:disable Metrics/MethodLength, Metrics/AbcSize
+      def unify_in_union(enum)
+        # We need to reset @unify before each branch is processed.
+        # Moreover we need to keep track of newly encountered wildcards.
+        # Var `new_unify_intersection` will hold those that are encountered
+        # in all branches; these are not a problem.
+        # Var `partial_unify` will hold those encountered in only a subset
+        # of the branches; these can't be used outside of the union.
+
+        return to_enum __method__, enum unless block_given?
+
+        new_unify_intersection = nil
+        partial_unify = []
+        unify_before = @unify.dup
+
+        result = enum.each do |e|
+          @unify = unify_before.dup if new_unify_intersection
+          yield e
+          new_unify = @unify.keys - unify_before.keys
+          if new_unify_intersection.nil?
+            # First iteration
+            new_unify_intersection = new_unify
+          else
+            union = new_unify_intersection | new_unify
+            new_unify_intersection &= new_unify
+            partial_unify |= union - new_unify_intersection
+          end
+        end
+
+        # At this point, all members of `new_unify_intersection` can be used
+        # for unification outside of the union, but partial_unify may not
+
+        forbid_unification(*partial_unify)
+
+        result
+      end
+      # rubocop:enable Metrics/MethodLength, Metrics/AbcSize
+
       def compile_union
         # we need to ensure that each branch of the {} contains the same
         # number of captures (since only one branch of the {} can actually
         # match, the same variables are used to hold the captures for each
         # branch)
         enum = tokens_until('}', 'union')
+        enum = unify_in_union(enum)
         terms = insure_same_captures(enum, 'branch of {}')
                 .map { compile_expr }
 
@@ -503,12 +558,12 @@ def compile_wildcard(name)
           # we have already seen a wildcard with this name before
           # so the value it matched the first time will already be stored
           # in a temp. check if this value matches the one stored in the temp
-          "#{CUR_ELEMENT} == temp#{@unify[name]}"
+          "#{CUR_ELEMENT} == #{access_unify(name)}"
         else
-          n = @unify[name] = next_temp_value
-          # double assign to temp#{n} to avoid "assigned but unused variable"
-          "(temp#{n} = #{CUR_ELEMENT}; " \
-          "temp#{n} = temp#{n}; true)"
+          n = @unify[name] = "unify_#{name.gsub('-', '__')}"
+          # double assign to avoid "assigned but unused variable"
+          "(#{n} = #{CUR_ELEMENT}; " \
+          "#{n} = #{n}; true)"
         end
       end
 
@@ -560,9 +615,8 @@ def compile_args(tokens)
       def compile_arg(token)
         case token
         when WILDCARD  then
-          name   = token[1..-1]
-          number = @unify[name] || fail_due_to('invalid in arglist: ' + token)
-          "temp#{number}"
+          name = token[1..-1]
+          access_unify(name) || fail_due_to('invalid in arglist: ' + token)
         when LITERAL   then token
         when PARAM     then get_param(token[1..-1])
         when CLOSING   then fail_due_to("#{token} in invalid position")
diff --git a/spec/rubocop/node_pattern_spec.rb b/spec/rubocop/node_pattern_spec.rb
index cd06147f8e3..8a09d769937 100644
--- a/spec/rubocop/node_pattern_spec.rb
+++ b/spec/rubocop/node_pattern_spec.rb
@@ -460,6 +460,110 @@
 
         it_behaves_like 'matching'
       end
+
+      context 'within a union' do
+        context 'confined to the union' do
+          context 'without unification' do
+            let(:pattern) { '{(array (int 1) _num) (array _num (int 1))}' }
+            let(:ruby) { '[2, 1]' }
+
+            it_behaves_like 'matching'
+          end
+
+          context 'with partial unification' do
+            let(:pattern) { '{(array _num _num) (array _num (int 1))}' }
+
+            context 'matching the unified branch' do
+              let(:ruby) { '[5, 5]' }
+
+              it_behaves_like 'matching'
+            end
+
+            context 'matching the free branch' do
+              let(:ruby) { '[2, 1]' }
+
+              it_behaves_like 'matching'
+            end
+
+            context 'that can not be unified' do
+              let(:ruby) { '[3, 2]' }
+
+              it_behaves_like 'nonmatching'
+            end
+          end
+        end
+
+        context 'with a preceeding unifying constraint' do
+          let(:pattern) do
+            '(array _num {(array (int 1) _num)
+                          send
+                          (array _num (int 1))})'
+          end
+
+          context 'matching a branch' do
+            let(:ruby) { '[2, [2, 1]]' }
+
+            it_behaves_like 'matching'
+          end
+
+          context 'that can not be unified' do
+            let(:ruby) { '[3, [2, 1]]' }
+
+            it_behaves_like 'nonmatching'
+          end
+        end
+
+        context 'with a succeeding unifying constraint' do
+          context 'with branches without the wildcard' do
+            context 'encountered first' do
+              let(:pattern) do
+                '(array {send
+                         (array (int 1) _num)
+                        } _num)'
+              end
+
+              it_behaves_like 'invalid'
+            end
+
+            context 'encountered after' do
+              let(:pattern) do
+                '(array {(array (int 1) _num)
+                         (array _num (int 1))
+                          send
+                        } _num)'
+              end
+
+              it_behaves_like 'invalid'
+            end
+          end
+
+          context 'with all branches with the wildcard' do
+            let(:pattern) do
+              '(array {(array (int 1) _num)
+                       (array _num (int 1))
+                      } _num)'
+            end
+
+            context 'matching the first branch' do
+              let(:ruby) { '[[1, 2], 2]' }
+
+              it_behaves_like 'matching'
+            end
+
+            context 'matching another branch' do
+              let(:ruby) { '[[2, 1], 2]' }
+
+              it_behaves_like 'matching'
+            end
+
+            context 'that can not be unified' do
+              let(:ruby) { '[[2, 1], 1]' }
+
+              it_behaves_like 'nonmatching'
+            end
+          end
+        end
+      end
     end
   end