Hi @danielavarelat,
Glad to hear that you were able to get past the DADA2 step!
Memory issues when classifying with Silva is a known problem. Some relatively recent discussion of this, and tips, are consolidated in this post by @Nicholas_Bokulich.
@Mehrbod_Estaki also suggested that filtering low abundance features might help at this stage. You could do that with qiime feature-table filter-features --p-min-samples 2 ...
(to include only features/ASVs that are present in at least two samples). You would do that filtering on your feature table, and then filter the features from your repseq.qza
file using qiime feature-table filter-seqs --i-table ...
. This type of filter can reduce the feature count by as much as half sometimes, which can help a bit with memory.
Another alternative would be to use a different reference database for classification, such as Greengenes2 (classifiers available here) or GTDB (see here for details on how to train one of those).