diff --git a/sqlchain/blockindex.go b/sqlchain/blockindex.go
index a5fd16dad..96d773bef 100644
--- a/sqlchain/blockindex.go
+++ b/sqlchain/blockindex.go
@@ -26,7 +26,7 @@ import (
 
 type blockNode struct {
 	parent *blockNode
-	block  *types.Block // TODO(leventeliu): cleanup history blocks to release memory.
+	block  *types.Block
 	hash   hash.Hash
 	height int32 // height is the chain height of the head
 	count  int32 // count counts the blocks (except genesis) at this head
diff --git a/sqlchain/chain.go b/sqlchain/chain.go
index 79e68e936..f2d79a851 100644
--- a/sqlchain/chain.go
+++ b/sqlchain/chain.go
@@ -44,6 +44,10 @@ import (
 	"github.com/syndtr/goleveldb/leveldb/util"
 )
 
+const (
+	minBlockCacheTTL = int32(100)
+)
+
 var (
 	metaState         = [4]byte{'S', 'T', 'A', 'T'}
 	metaBlockIndex    = [4]byte{'B', 'L', 'C', 'K'}
@@ -359,6 +363,7 @@ func LoadChain(c *Config) (chain *Chain, err error) {
 	st.node = last
 	chain.rt.setHead(st)
 	chain.st.InitTx(id)
+	chain.pruneBlockCache()
 
 	// Read queries and rebuild memory index
 	respIter := chain.tdb.NewIterator(util.BytesPrefix(metaResponseIndex[:]), nil)
@@ -695,6 +700,7 @@ func (c *Chain) syncHead() {
 // runCurrentTurn does the check and runs block producing if its my turn.
 func (c *Chain) runCurrentTurn(now time.Time) {
 	defer func() {
+		c.pruneBlockCache()
 		c.rt.setNextTurn()
 		c.qi.advanceBarrier(c.rt.getMinValidHeight())
 		c.ai.advance(c.rt.getMinValidHeight())
@@ -1187,6 +1193,11 @@ func (c *Chain) getBilling(low, high int32) (req *pt.BillingRequest, err error)
 	}
 
 	for ; n != nil && n.height >= low; n = n.parent {
+		// TODO(leventeliu): block maybe released, use persistence version in this case.
+		if n.block == nil {
+			continue
+		}
+
 		if lowBlock == nil {
 			lowBlock = n.block
 		}
@@ -1521,3 +1532,21 @@ func (c *Chain) register(ack *types.SignedAckHeader) (err error) {
 func (c *Chain) remove(ack *types.SignedAckHeader) (err error) {
 	return c.ai.remove(c.rt.getHeightFromTime(ack.SignedRequestHeader().Timestamp), ack)
 }
+
+func (c *Chain) pruneBlockCache() {
+	var (
+		head    = c.rt.getHead().node
+		lastCnt int32
+	)
+	if head == nil {
+		return
+	}
+	lastCnt = head.count - c.rt.blockCacheTTL
+	// Move to last count position
+	for ; head != nil && head.count > lastCnt; head = head.parent {
+	}
+	// Prune block references
+	for ; head != nil && head.block != nil; head = head.parent {
+		head.block = nil
+	}
+}
diff --git a/sqlchain/config.go b/sqlchain/config.go
index 9fad34d91..113e99718 100644
--- a/sqlchain/config.go
+++ b/sqlchain/config.go
@@ -46,6 +46,8 @@ type Config struct {
 	// QueryTTL sets the unacknowledged query TTL in block periods.
 	QueryTTL int32
 
+	BlockCacheTTL int32
+
 	// DBAccount info
 	TokenType    pt.TokenType
 	GasPrice     uint64
diff --git a/sqlchain/runtime.go b/sqlchain/runtime.go
index 1537452dc..b1935558d 100644
--- a/sqlchain/runtime.go
+++ b/sqlchain/runtime.go
@@ -47,6 +47,8 @@ type runtime struct {
 	tick time.Duration
 	// queryTTL sets the unacknowledged query TTL in block periods.
 	queryTTL int32
+	// blockCacheTTL sets the cached block numbers.
+	blockCacheTTL int32
 	// muxServer is the multiplexing service of sql-chain PRC.
 	muxService *MuxService
 	// price sets query price in gases.
@@ -85,11 +87,17 @@ type runtime struct {
 // newRunTime returns a new sql-chain runtime instance with the specified config.
 func newRunTime(c *Config) (r *runtime) {
 	r = &runtime{
-		stopCh:          make(chan struct{}),
-		databaseID:      c.DatabaseID,
-		period:          c.Period,
-		tick:            c.Tick,
-		queryTTL:        c.QueryTTL,
+		stopCh:     make(chan struct{}),
+		databaseID: c.DatabaseID,
+		period:     c.Period,
+		tick:       c.Tick,
+		queryTTL:   c.QueryTTL,
+		blockCacheTTL: func() int32 {
+			if c.BlockCacheTTL < minBlockCacheTTL {
+				return minBlockCacheTTL
+			}
+			return c.BlockCacheTTL
+		}(),
 		muxService:      c.MuxService,
 		price:           c.Price,
 		producingReward: c.ProducingReward,