diff --git a/src/metric.cpp b/src/metric.cpp
index 836360b..bb278b5 100644
--- a/src/metric.cpp
+++ b/src/metric.cpp
@@ -244,7 +244,16 @@ template <> struct Family<Counter>::State {
   std::string_view help;
 
   struct PerThreadState {
-    std::unordered_map<LabelsKey, Counter::State *> instances;
+    std::unordered_map<
+        LabelsKey, Counter::State *, std::hash<LabelsKey>,
+        std::equal_to<LabelsKey>,
+        ArenaStlAllocator<std::pair<const LabelsKey, Counter::State *>>>
+        instances;
+
+    explicit PerThreadState(ArenaAllocator &arena)
+        : instances(
+              ArenaStlAllocator<std::pair<const LabelsKey, Counter::State *>>(
+                  &arena)) {}
   };
   std::unordered_map<std::thread::id, PerThreadState> per_thread_state;
 
@@ -296,7 +305,16 @@ template <> struct Family<Histogram>::State {
   ArenaVector<double> buckets;
 
   struct PerThreadState {
-    std::unordered_map<LabelsKey, Histogram::State *> instances;
+    std::unordered_map<
+        LabelsKey, Histogram::State *, std::hash<LabelsKey>,
+        std::equal_to<LabelsKey>,
+        ArenaStlAllocator<std::pair<const LabelsKey, Histogram::State *>>>
+        instances;
+
+    explicit PerThreadState(ArenaAllocator &arena)
+        : instances(
+              ArenaStlAllocator<std::pair<const LabelsKey, Histogram::State *>>(
+                  &arena)) {}
   };
   std::unordered_map<std::thread::id, PerThreadState> per_thread_state;
 
@@ -588,9 +606,16 @@ struct Metric {
 
     // Ensure thread state exists
     auto thread_id = std::this_thread::get_id();
-    // Thread state is automatically created by operator[]
+    auto per_thread_it = family->p->per_thread_state.find(thread_id);
+    if (per_thread_it == family->p->per_thread_state.end()) {
+      // Create new PerThreadState with thread-local arena
+      auto result = family->p->per_thread_state.emplace(
+          thread_id,
+          Family<Counter>::State::PerThreadState(get_thread_local_arena()));
+      per_thread_it = result.first;
+    }
 
-    auto &ptr = family->p->per_thread_state[thread_id].instances[key];
+    auto &ptr = per_thread_it->second.instances[key];
     if (!ptr) {
       ptr = get_thread_local_arena().construct<Counter::State>();
       ptr->value = 0.0;
@@ -642,9 +667,16 @@ struct Metric {
 
     // Ensure thread state exists
     auto thread_id = std::this_thread::get_id();
-    // Thread state is automatically created by operator[]
+    auto per_thread_it = family->p->per_thread_state.find(thread_id);
+    if (per_thread_it == family->p->per_thread_state.end()) {
+      // Create new PerThreadState with thread-local arena
+      auto result = family->p->per_thread_state.emplace(
+          thread_id,
+          Family<Histogram>::State::PerThreadState(get_thread_local_arena()));
+      per_thread_it = result.first;
+    }
 
-    auto &ptr = family->p->per_thread_state[thread_id].instances[key];
+    auto &ptr = per_thread_it->second.instances[key];
     if (!ptr) {
       ptr = get_thread_local_arena().construct<Histogram::State>();