diff --git a/.changeset/fix-mmseqs2-memory-headroom.md b/.changeset/fix-mmseqs2-memory-headroom.md
new file mode 100644
index 0000000..c4437c4
--- /dev/null
+++ b/.changeset/fix-mmseqs2-memory-headroom.md
@@ -0,0 +1,5 @@
+---
+'@platforma-open/milaboratories.clonotype-clustering.workflow': patch
+---
+
+Reduce MMseqs2 --split-memory-limit from 80% to 60% of system RAM to prevent OOM kills
diff --git a/workflow/src/clustering.tpl.tengo b/workflow/src/clustering.tpl.tengo
index 973b060..a768831 100644
--- a/workflow/src/clustering.tpl.tengo
+++ b/workflow/src/clustering.tpl.tengo
@@ -126,7 +126,10 @@ self.body(func(inputs) {
             run()
 
         // get split memory limit based on real system ram
-	    memLimit := "{int(ceil(system.ram.gb * 0.8))}" + "G"
+        // --split-memory-limit controls only the prefilter database splitting;
+        // actual peak usage is ~125% of this value, plus uncontrolled memory
+        // from clustering and alignment steps; 60% leaves headroom for all of that
+	    memLimit := "{int(max(system.ram.gb*60/100,1))}" + "G"
 
         mem := "32GiB" // @TODO: set based on the size of the input
 	    cpu := 16