When I read the capture file mentioned in bug 12367
https://bugs.wireshark.org/bugzilla/show_bug.cgi?id=12367
it eats about 6-8GB on my machine.
A large amount of that data is in structures allocated by init_tcp_conversation_data(), which is called by get_tcp_conversation_data() if there isn't already one for the conversation.
get_tcp_conversation_data() is *always* called by dissect_tcp(), so you can't disable that analysis.
So if you're reading a large capture file with a lot of TCP connections, make sure you're on a 64-bit machine that has plenty of memory and that either has or can allocate plenty of swap space to back it if necessary.