Improving the memory density of in-memory columnar databases is a necessary step towards more efficient sharing of hardware in multi-tenant systems. Many general-purpose tools attempt to achieve this by scanning systems for duplicate memory regions, and deduplicating these redundancies. In this paper, we show that existing tools sacrifice efficiency for general-applicability, and outline the possibilities unlocked by tapping into the architectural guarantees, and existing metadata for in-memory databases. Based on our previous work on deduplication, where we used a hardware-accelerated delta mechanism to increase memory savings, we detail a process for semantics-based steering of this mechanism. This significantly reduces the search space for duplicates, by only comparing database columns which exhibit similar features. Furthermore, we gracefully handle cases where steering causes the comparison of mismatched columns. We benchmark our solution against the widely available Linux Kernel Samepage Merging (KSM) tool, evaluating both based on their achieved memory savings and execution speed for deduplicating the Internet Movie Data Base (IMDB) and Star Schema Benchmark (SSB) datasets. Our solution manages to perform an order of magnitude faster than KSM, while detecting close to the same amount of duplicate data.