diff --git a/docs/CHANGELOG.md b/docs/CHANGELOG.md
index 20e7131478..ea5dd36f59 100644
--- a/docs/CHANGELOG.md
+++ b/docs/CHANGELOG.md
@@ -61,6 +61,7 @@ The sandbox cluster installation is running under the constant load generated by
 * BUGFIX: [vmctl](https://docs.victoriametrics.com/vmctl.html): retry on import errors in `vm-native` mode. Before, retries happened only on writes into a network connection between source and destination. But errors returned by server after all the data was transmitted were logged, but not retried.
 * BUGFIX: [vmagent](https://docs.victoriametrics.com/vmagent.html): properly assume role with [AWS IRSA authorization](https://docs.aws.amazon.com/eks/latest/userguide/iam-roles-for-service-accounts.html). Previously role chaining was not supported. See [this issue](https://github.com/VictoriaMetrics/VictoriaMetrics/issues/3822) for details.
 * BUGFIX: [vmagent](https://docs.victoriametrics.com/vmagent.html): exit if there is config syntax error in [`scrape_config_files`](https://docs.victoriametrics.com/vmagent.html#loading-scrape-configs-from-multiple-files) when `-promscrape.config.strictParse=true`. See [this issue](https://github.com/VictoriaMetrics/VictoriaMetrics/issues/5508).
+* BUGFIX: [vmagent](https://docs.victoriametrics.com/vmagent.html): do not store scrape response for target in memory when staleness markers are disabled. See [this pull request](https://github.com/VictoriaMetrics/VictoriaMetrics/pull/5577) for details.
 * BUGFIX: [vmui](https://docs.victoriametrics.com/#vmui): fix a link for the statistic inaccuracy explanation in the cardinality explorer tool. See [this issue](https://github.com/VictoriaMetrics/VictoriaMetrics/issues/5460).
 * BUGFIX: all: fix potential panic during components shutdown when [metrics push](https://docs.victoriametrics.com/#push-metrics) is configured. See [this issue](https://github.com/VictoriaMetrics/VictoriaMetrics/issues/5548). Thanks to @zhdd99 for the [pull request](https://github.com/VictoriaMetrics/VictoriaMetrics/pull/5549).
 * BUGFIX: [vmselect](https://docs.victoriametrics.com/Cluster-VictoriaMetrics.html): properly determine time range search for instant queries with too big look-behind window like `foo[100y]`. Previously, such queries could return empty responses even if `foo` is present in database.
diff --git a/lib/promscrape/scrapework.go b/lib/promscrape/scrapework.go
index 5858ae8801..a5fa45aa53 100644
--- a/lib/promscrape/scrapework.go
+++ b/lib/promscrape/scrapework.go
@@ -337,11 +337,13 @@ func (sw *scrapeWork) run(stopCh <-chan struct{}, globalStopCh <-chan struct{})
 				// Do not send staleness markers on graceful shutdown as Prometheus does.
 				// See https://github.com/VictoriaMetrics/VictoriaMetrics/issues/2013#issuecomment-1006994079
 			default:
-				// Send staleness markers to all the metrics scraped last time from the target
-				// when the given target disappears as Prometheus does.
-				// Use the current real timestamp for staleness markers, so queries
-				// stop returning data just after the time the target disappears.
-				sw.sendStaleSeries(lastScrape, "", t, true)
+				if !sw.Config.NoStaleMarkers {
+					// Send staleness markers to all the metrics scraped last time from the target
+					// when the given target disappears as Prometheus does.
+					// Use the current real timestamp for staleness markers, so queries
+					// stop returning data just after the time the target disappears.
+					sw.sendStaleSeries(lastScrape, "", t, true)
+				}
 			}
 			if sw.seriesLimiter != nil {
 				sw.seriesLimiter.MustStop()
@@ -487,7 +489,8 @@ func (sw *scrapeWork) processScrapedData(scrapeTimestamp, realTimestamp int64, b
 		bodyString = ""
 	}
 	seriesAdded := 0
-	if !areIdenticalSeries {
+	// cannot track the number of new time series when stale markers are disabled.
+	if !sw.Config.NoStaleMarkers && !areIdenticalSeries {
 		// The returned value for seriesAdded may be bigger than the real number of added series
 		// if some series were removed during relabeling.
 		// This is a trade-off between performance and accuracy.
@@ -517,7 +520,7 @@ func (sw *scrapeWork) processScrapedData(scrapeTimestamp, realTimestamp int64, b
 		writeRequestCtxPool.Put(wc)
 	}
 	// body must be released only after wc is released, since wc refers to body.
-	if !areIdenticalSeries {
+	if !sw.Config.NoStaleMarkers && !areIdenticalSeries {
 		// Send stale markers for disappeared metrics with the real scrape timestamp
 		// in order to guarantee that query doesn't return data after this time for the disappeared metrics.
 		sw.sendStaleSeries(lastScrape, bodyString, realTimestamp, false)
@@ -627,7 +630,8 @@ func (sw *scrapeWork) scrapeStream(scrapeTimestamp, realTimestamp int64) error {
 		scrapesFailed.Inc()
 	}
 	seriesAdded := 0
-	if !areIdenticalSeries {
+	// cannot track the number of new time series when stale markers are disabled.
+	if !sw.Config.NoStaleMarkers && !areIdenticalSeries {
 		// The returned value for seriesAdded may be bigger than the real number of added series
 		// if some series were removed during relabeling.
 		// This is a trade-off between performance and accuracy.
@@ -647,7 +651,7 @@ func (sw *scrapeWork) scrapeStream(scrapeTimestamp, realTimestamp int64) error {
 	sw.prevBodyLen = sbr.bodyLen
 	wc.reset()
 	writeRequestCtxPool.Put(wc)
-	if !areIdenticalSeries {
+	if !sw.Config.NoStaleMarkers && !areIdenticalSeries {
 		// Send stale markers for disappeared metrics with the real scrape timestamp
 		// in order to guarantee that query doesn't return data after this time for the disappeared metrics.
 		sw.sendStaleSeries(lastScrape, bodyString, realTimestamp, false)
@@ -788,9 +792,6 @@ func (sw *scrapeWork) sendStaleSeries(lastScrape, currScrape string, timestamp i
 	defer func() {
 		<-sendStaleSeriesConcurrencyLimitCh
 	}()
-	if sw.Config.NoStaleMarkers {
-		return
-	}
 	bodyString := lastScrape
 	if currScrape != "" {
 		bodyString = parser.GetRowsDiff(lastScrape, currScrape)