diff --git a/app/views/home/_stats.html.haml b/app/views/home/_stats.html.haml
index 7ad0df6ae..949a98008 100644
--- a/app/views/home/_stats.html.haml
+++ b/app/views/home/_stats.html.haml
@@ -1,4 +1,6 @@
-- cache("homepage_stats") do
+-# expires_in bounds staleness: expire_fragment only clears the cache of the
+-# process that handled the write, and the cache store may be per process.
+- cache("homepage_stats", expires_in: 12.hours) do
   %p.stats
     = t('.message_html',
      member: link_to(t('.member_linktext', count: Member.confirmed.size.to_i), members_path),
diff --git a/config/environments/production.rb b/config/environments/production.rb
index 9bfa08117..4caba1b9e 100644
--- a/config/environments/production.rb
+++ b/config/environments/production.rb
@@ -51,15 +51,22 @@ Rails.application.configure do
   # Prepend all log lines with the following tags.
   config.log_tags = [:request_id]
 
-  # Use a different cache store in production.
-  config.cache_store = :mem_cache_store,
-                       (ENV["MEMCACHIER_SERVERS"] || "").split(","),
-                       { username:             ENV.fetch("MEMCACHIER_USERNAME", nil),
-                         password:             ENV.fetch("MEMCACHIER_PASSWORD", nil),
-                         failover:             true,
-                         socket_timeout:       0.5,
-                         socket_failure_delay: 0.2,
-                         down_retry_delay:     60 }
+  # Memcached is opt-in via USE_MEMCACHED. Otherwise each process keeps its own
+  # capped in-memory cache: the store is per Puma worker, so the dyno pays
+  # RAILS_CACHE_SIZE_MB once per worker, not once in total.
+  memcached_servers = ENV.fetch("MEMCACHIER_SERVERS", "").split(",")
+
+  config.cache_store = if ENV["USE_MEMCACHED"] == "true" && memcached_servers.any?
+                         [:mem_cache_store, memcached_servers,
+                          { username:             ENV.fetch("MEMCACHIER_USERNAME", nil),
+                            password:             ENV.fetch("MEMCACHIER_PASSWORD", nil),
+                            failover:             true,
+                            socket_timeout:       0.5,
+                            socket_failure_delay: 0.2,
+                            down_retry_delay:     60 }]
+                       else
+                         [:memory_store, { size: Integer(ENV.fetch("RAILS_CACHE_SIZE_MB", "32")).megabytes }]
+                       end
 
   # Use a real queuing backend for Active Job (and separate queues per environment)
   # config.active_job.queue_adapter     = :resque