...
AEM 6.x with Apache Sling uses TarMK with Apache Jackrabbit Oak. Key aspect of speed is the combination deserializes the repository memory. It is not caching the repository in memory. The repository exists in memory and uses the tar files to serialize adn persist the data. It will deserialize the entire repository into memory if enough memory available. So give AEM lots of memory.
Though AEM and Apache Sling runs in a JVM, the memory they use for repository is not the JVM heap, but the "off-heap" memory. You may limit it via Java parameter -XX:MaxDirectMemorySize. One consideration to gain memory is to set the Pool Size parameter of Apache Sling Job Thread Pool to 1. This iwll limit the memory used by jobs at the cost of speed of job completion when jobs are queued.
Setup of Web Server Dispatcher
...