diff --git a/tsdb/chunkenc/chunk.go b/tsdb/chunkenc/chunk.go
index 5f9349f05..5f228df48 100644
--- a/tsdb/chunkenc/chunk.go
+++ b/tsdb/chunkenc/chunk.go
@@ -41,14 +41,29 @@ const (
 
 // Chunk holds a sequence of sample pairs that can be iterated over and appended to.
 type Chunk interface {
+	// Bytes returns the underlying byte slice of the chunk.
 	Bytes() []byte
+
+	// Encoding returns the encoding type of the chunk.
 	Encoding() Encoding
+
+	// Appender returns an appender to append samples to the chunk.
 	Appender() (Appender, error)
+
 	// The iterator passed as argument is for re-use.
 	// Depending on implementation, the iterator can
 	// be re-used or a new iterator can be allocated.
 	Iterator(Iterator) Iterator
+
+	// NumSamples returns the number of samples in the chunk.
 	NumSamples() int
+
+	// Compact is called whenever a chunk is expected to be complete (no more
+	// samples appended) and the underlying implementation can eventually
+	// optimize the chunk.
+	// There's no strong guarantee that no samples will be appended once
+	// Compact() is called. Implementing this function is optional.
+	Compact()
 }
 
 // Appender adds sample pairs to a chunk.
diff --git a/tsdb/chunkenc/xor.go b/tsdb/chunkenc/xor.go
index ce6e0a951..5401261d9 100644
--- a/tsdb/chunkenc/xor.go
+++ b/tsdb/chunkenc/xor.go
@@ -49,6 +49,10 @@ import (
 	"math/bits"
 )
 
+const (
+	chunkCompactCapacityThreshold = 32
+)
+
 // XORChunk holds XOR encoded sample data.
 type XORChunk struct {
 	b bstream
@@ -75,6 +79,14 @@ func (c *XORChunk) NumSamples() int {
 	return int(binary.BigEndian.Uint16(c.Bytes()))
 }
 
+func (c *XORChunk) Compact() {
+	if l := len(c.b.stream); cap(c.b.stream) > l+chunkCompactCapacityThreshold {
+		buf := make([]byte, l)
+		copy(buf, c.b.stream)
+		c.b.stream = buf
+	}
+}
+
 // Appender implements the Chunk interface.
 func (c *XORChunk) Appender() (Appender, error) {
 	it := c.iterator(nil)
diff --git a/tsdb/head.go b/tsdb/head.go
index ce4a36719..5769c770a 100644
--- a/tsdb/head.go
+++ b/tsdb/head.go
@@ -1691,6 +1691,11 @@ func (s *memSeries) cut(mint int64) *memChunk {
 	s.chunks = append(s.chunks, c)
 	s.headChunk = c
 
+	// Remove exceeding capacity from the previous chunk byte slice to save memory.
+	if l := len(s.chunks); l > 1 {
+		s.chunks[l-2].chunk.Compact()
+	}
+
 	// Set upper bound on when the next chunk must be started. An earlier timestamp
 	// may be chosen dynamically at a later point.
 	s.nextAt = rangeForTimestamp(mint, s.chunkRange)