In non-volatile memory devices, the accessing of data on word line can degrade the data quality on a neighboring word line, in what is called a read disturb. Techniques are presented for determining word lines likely to suffer read disturbs by use of a hash tree for tracking the number of reads. Read counters are maintained for memory units at a relatively coarse granularity, such as a die or block. When the counter for one of these units reaches a certain level, it is subdivided into sub-units, each with their own read counter, in a process that be repeated to determine frequently read word lines with a fine level of granularity while only using a relatively modest amount of RAM on the controller to store the counters.