YCM uses racerd (https://github.com/jwilm/racerd) which keeps everything in a daemon process's memory instead of reparsing every time. It's a good bit faster (at least on my machine), but it uses racer under the hood for recommendations.