diff --git a/promql/engine.go b/promql/engine.go
index edc90d21a2..43a47c6cc7 100644
--- a/promql/engine.go
+++ b/promql/engine.go
@@ -1016,6 +1016,7 @@ type evaluator struct {
 	lookbackDelta            time.Duration
 	samplesStats             *stats.QuerySamples
 	noStepSubqueryIntervalFn func(rangeMillis int64) int64
+	reuseHistograms          bool
 }
 
 // errorf causes a panic with the input formatted into an error.
@@ -2152,10 +2153,19 @@ loop:
 					histograms = getMatrixSelectorHPoints()
 				}
 				n := len(histograms)
-				if n < cap(histograms) {
-					histograms = histograms[:n+1]
+				if ev.reuseHistograms {
+					if n < cap(histograms) {
+						histograms = histograms[:n+1]
+					} else {
+						histograms = append(histograms, HPoint{H: &histogram.FloatHistogram{}})
+					}
 				} else {
-					histograms = append(histograms, HPoint{H: &histogram.FloatHistogram{}})
+					if n < cap(histograms) {
+						histograms = histograms[:n+1]
+						histograms[n].H = nil
+					} else {
+						histograms = append(histograms, HPoint{})
+					}
 				}
 				histograms[n].T, histograms[n].H = buf.AtFloatHistogram(histograms[n].H)
 				if value.IsStaleNaN(histograms[n].H.Sum) {
@@ -2195,10 +2205,19 @@ loop:
 			histograms = getMatrixSelectorHPoints()
 		}
 		n := len(histograms)
-		if n < cap(histograms) {
-			histograms = histograms[:n+1]
+		if ev.reuseHistograms {
+			if n < cap(histograms) {
+				histograms = histograms[:n+1]
+			} else {
+				histograms = append(histograms, HPoint{H: &histogram.FloatHistogram{}})
+			}
 		} else {
-			histograms = append(histograms, HPoint{H: &histogram.FloatHistogram{}})
+			if n < cap(histograms) {
+				histograms = histograms[:n+1]
+				histograms[n].H = nil
+			} else {
+				histograms = append(histograms, HPoint{})
+			}
 		}
 		histograms[n].T, histograms[n].H = it.AtFloatHistogram(histograms[n].H)
 		if value.IsStaleNaN(histograms[n].H.Sum) {
diff --git a/storage/buffer.go b/storage/buffer.go
index b3c789e97d..47c85a5ad0 100644
--- a/storage/buffer.go
+++ b/storage/buffer.go
@@ -240,6 +240,8 @@ func (s fhSample) Type() chunkenc.ValueType {
 type sampleRing struct {
 	delta int64
 
+	reuseHistograms bool
+
 	// Lookback buffers. We use iBuf for mixed samples, but one of the three
 	// concrete ones for homogenous samples. (Only one of the four bufs is
 	// allowed to be populated!) This avoids the overhead of the interface
@@ -385,7 +387,7 @@ func (it *SampleRingIterator) AtFloatHistogram(fh *histogram.FloatHistogram) (in
 		it.fh.CopyTo(fh)
 		return it.t, fh
 	}
-	return it.t, it.fh.Copy()
+	return it.t, it.fh
 }
 
 func (it *SampleRingIterator) AtT() int64 {
@@ -680,10 +682,14 @@ func addH(s hSample, buf []hSample, r *sampleRing) []hSample {
 	}
 
 	buf[r.i].t = s.t
-	if buf[r.i].h == nil {
-		buf[r.i].h = s.h.Copy()
+	if r.reuseHistograms {
+		if buf[r.i].h == nil {
+			buf[r.i].h = s.h.Copy()
+		} else {
+			s.h.CopyTo(buf[r.i].h)
+		}
 	} else {
-		s.h.CopyTo(buf[r.i].h)
+		buf[r.i].h = s.h
 	}
 	r.l++
 
@@ -724,10 +730,14 @@ func addFH(s fhSample, buf []fhSample, r *sampleRing) []fhSample {
 	}
 
 	buf[r.i].t = s.t
-	if buf[r.i].fh == nil {
-		buf[r.i].fh = s.fh.Copy()
+	if r.reuseHistograms {
+		if buf[r.i].fh == nil {
+			buf[r.i].fh = s.fh.Copy()
+		} else {
+			s.fh.CopyTo(buf[r.i].fh)
+		}
 	} else {
-		s.fh.CopyTo(buf[r.i].fh)
+		buf[r.i].fh = s.fh
 	}
 	r.l++