diff --git a/bwa.c b/bwa.c
index aff9aff..c8400b1 100644
--- a/bwa.c
+++ b/bwa.c
@@ -25,7 +25,7 @@ static inline void trim_readno(kstring_t *s)
 static inline void kseq2bseq1(const kseq_t *ks, bseq1_t *s)
 { // TODO: it would be better to allocate one chunk of memory, but probably it does not matter in practice
 	s->name = strdup(ks->name.s);
-	s->comment = ks->comment.l? strdup(s->comment) : 0;
+	s->comment = ks->comment.l? strdup(ks->comment.s) : 0;
 	s->seq = strdup(ks->seq.s);
 	s->qual = ks->qual.l? strdup(ks->qual.s) : 0;
 	s->l_seq = strlen(s->seq);
diff --git a/bwamem.c b/bwamem.c
index ce55cad..553fe1c 100644
--- a/bwamem.c
+++ b/bwamem.c
@@ -13,8 +13,6 @@
 #include "kvec.h"
 #include "ksort.h"
 
-int mem_verbose = 3; // 1: error only; 2: error+warning; 3: message+error+warning; >=4: debugging
-
 void mem_fill_scmat(int a, int b, int8_t mat[25])
 {
 	int i, j, k;
@@ -462,7 +460,7 @@ void mem_chain2aln(const mem_opt_t *opt, int64_t l_pac, const uint8_t *pac, int
 			a->score = ksw_extend(l_query - qe, query + qe, rmax[1] - rmax[0] - re, rseq + re, 5, opt->mat, opt->q, opt->r, opt->w, a->score, &qle, &tle);
 			a->qe = qe + qle; a->re = rmax[0] + re + tle;
 		} else a->qe = l_query, a->re = s->rbeg + s->len;
-		if (mem_verbose >= 4) printf("[%d] score=%d\t[%d,%d) <=> [%ld,%ld)\n", k, a->score, a->qb, a->qe, (long)a->rb, (long)a->re);
+		if (bwa_verbose >= 4) printf("[%d] score=%d\t[%d,%d) <=> [%ld,%ld)\n", k, a->score, a->qb, a->qe, (long)a->rb, (long)a->re);
 		// compute seedcov
 		for (i = 0, a->seedcov = 0; i < c->n; ++i) {
 			const mem_seed_t *t = &c->seeds[i];
@@ -572,6 +570,7 @@ void bwa_hit2sam(kstring_t *str, const int8_t mat[25], int q, int r, int w, cons
 	if (p->score >= 0) { kputsn("\tAS:i:", 6, str); kputw(p->score, str); }
 	if (p->sub >= 0) { kputsn("\tXS:i:", 6, str); kputw(p->sub, str); }
 	if (bwa_rg_id) { kputsn("\tRG:i:", 6, str); kputs(bwa_rg_id, str); }
+	if (s->comment) { kputc('\t', str); kputs(s->comment, str); }
 	kputc('\n', str);
 	free(cigar);
 #undef is_mapped
@@ -633,7 +632,7 @@ static mem_alnreg_v find_alnreg(const mem_opt_t *opt, const bwt_t *bwt, const bn
 		s->seq[i] = nst_nt4_table[(int)s->seq[i]];
 	chn = mem_chain(opt, bwt, s->l_seq, (uint8_t*)s->seq);
 	chn.n = mem_chain_flt(opt, chn.n, chn.a);
-	if (mem_verbose >= 4) mem_print_chain(bns, &chn);
+	if (bwa_verbose >= 4) mem_print_chain(bns, &chn);
 	kv_init(regs); kv_init(tmp);
 	for (i = 0; i < chn.n; ++i) {
 		mem_chain2aln(opt, bns->l_pac, pac, s->l_seq, (uint8_t*)s->seq, &chn.a[i], &tmp);
diff --git a/bwamem_pair.c b/bwamem_pair.c
index 3dce119..57a128a 100644
--- a/bwamem_pair.c
+++ b/bwamem_pair.c
@@ -47,7 +47,7 @@ void mem_pestat(const mem_opt_t *opt, int64_t l_pac, int n, const mem_alnreg_v *
 		dir = mem_infer_dir(l_pac, r[0]->a[0].rb, r[1]->a[0].rb, &is);
 		if (is && is <= opt->max_ins) kv_push(uint64_t, isize[dir], is);
 	}
-	if (mem_verbose >= 3) fprintf(stderr, "[M::%s] # candidate unique pairs for (FF, FR, RF, RR): (%ld, %ld, %ld, %ld)\n", __func__, isize[0].n, isize[1].n, isize[2].n, isize[3].n);
+	if (bwa_verbose >= 3) fprintf(stderr, "[M::%s] # candidate unique pairs for (FF, FR, RF, RR): (%ld, %ld, %ld, %ld)\n", __func__, isize[0].n, isize[1].n, isize[2].n, isize[3].n);
 	for (d = 0; d < 4; ++d) { // TODO: this block is nearly identical to the one in bwtsw2_pair.c. It would be better to merge these two.
 		mem_pestat_t *r = &pes[d];
 		uint64_v *q = &isize[d];
diff --git a/fastmap.c b/fastmap.c
index adbe04c..437192a 100644
--- a/fastmap.c
+++ b/fastmap.c
@@ -14,7 +14,7 @@ extern unsigned char nst_nt4_table[256];
 int main_mem(int argc, char *argv[])
 {
 	mem_opt_t *opt;
-	int c, n;
+	int i, c, n, copy_comment = 0;
 	gzFile fp, fp2 = 0;
 	kseq_t *ks, *ks2 = 0;
 	bseq1_t *seqs;
@@ -22,14 +22,15 @@ int main_mem(int argc, char *argv[])
 	char *rg_line = 0;
 
 	opt = mem_opt_init();
-	while ((c = getopt(argc, argv, "PHk:c:v:s:r:t:R:")) >= 0) {
+	while ((c = getopt(argc, argv, "CPHk:c:v:s:r:t:R:")) >= 0) {
 		if (c == 'k') opt->min_seed_len = atoi(optarg);
 		else if (c == 't') opt->n_threads = atoi(optarg), opt->n_threads = opt->n_threads > 1? opt->n_threads : 1;
 		else if (c == 'P') opt->flag |= MEM_F_NOPAIRING;
 		else if (c == 'H') opt->flag |= MEM_F_HARDCLIP;
 		else if (c == 'c') opt->max_occ = atoi(optarg);
-		else if (c == 'v') mem_verbose = atoi(optarg);
+		else if (c == 'v') bwa_verbose = atoi(optarg);
 		else if (c == 'r') opt->split_factor = atof(optarg);
+		else if (c == 'C') copy_comment = 1;
 		else if (c == 'R') {
 			if ((rg_line = bwa_set_rg(optarg)) == 0) return 1; // FIXME: memory leak
 		} else if (c == 's') opt->split_width = atoi(optarg);
@@ -43,7 +44,8 @@ int main_mem(int argc, char *argv[])
 		fprintf(stderr, "         -s INT     look for internal seeds inside a seed with less than INT occ [%d]\n", opt->split_width);
 		fprintf(stderr, "         -r FLOAT   look for internal seeds inside a seed longer than {-k} * FLOAT [%g]\n", opt->split_factor);
 		fprintf(stderr, "         -R STR     read group header line such as '@RG\tID:foo\tSM:bar' [null]\n");
-		fprintf(stderr, "         -v INT     verbose level [%d]\n", mem_verbose);
+		fprintf(stderr, "         -v INT     verbose level [%d]\n", bwa_verbose);
+		fprintf(stderr, "         -C         append FASTA/FASTQ comment to SAM output\n");
 		fprintf(stderr, "\n");
 		free(opt);
 		return 1;
@@ -60,6 +62,14 @@ int main_mem(int argc, char *argv[])
 		opt->flag |= MEM_F_PE;
 	}
 	while ((seqs = bseq_read(opt->chunk_size, &n, ks, ks2)) != 0) {
+		int64_t size = 0;
+		if (!copy_comment)
+			for (i = 0; i < n; ++i) {
+				free(seqs[i].comment); seqs[i].comment = 0;
+			}
+		for (i = 0; i < n; ++i) size += seqs[i].l_seq;
+		if (bwa_verbose >= 3)
+			fprintf(stderr, "[M::%s] read %d sequences (%ld bp)...\n", __func__, n, (long)size);
 		mem_process_seqs(opt, idx->bwt, idx->bns, idx->pac, n, seqs);
 		free(seqs);
 	}