diff --git a/lib/mu-container.c b/lib/mu-container.c
index 59633032..64f9bf9a 100644
--- a/lib/mu-container.c
+++ b/lib/mu-container.c
@@ -80,15 +80,6 @@ set_parent (MuContainer *c, MuContainer *parent)
 	}
 }
 
-static MuContainer*
-find_last (MuContainer *c)
-{
-	while (c && c->next)
-		c = c->next;
-
-	return c;
-}
-
 
 G_GNUC_UNUSED static gboolean
 check_dup (MuContainer *c, GHashTable *hash)
@@ -131,7 +122,21 @@ mu_container_append_siblings (MuContainer *c, MuContainer *sibling)
 	/* assert_no_duplicates (c); */
 
 	set_parent (sibling, c->parent);
-	(find_last(c))->next = sibling;
+
+	/* find the last sibling and append; first we try our cache
+	 * 'last', otherwise we need to walk the chain. We use a
+	 * cached last as to avoid walking the chain (which is
+	 * O(n*n)) */
+	if (c->last)
+		c->last->next = sibling;
+	else {
+		/* no 'last' cached, so walk the chain */
+		MuContainer *c2;
+		for (c2 = c; c2 && c2->next; c2 = c2->next);
+		c2->next = sibling;
+	}
+	/* update the cached last */
+	c->last = sibling->last ? sibling->last : sibling;
 
 	/* assert_no_duplicates (c); */
 
@@ -158,6 +163,13 @@ mu_container_remove_sibling (MuContainer *c, MuContainer *sibling)
 		prev = cur;
 	}
 
+	/* unset the cached last; it's not valid anymore
+	 *
+	 * TODO: we could actually do a better job updating last
+	 * rather than invalidating it. */
+	if (c)
+		c->last = NULL;
+
 	return c;
 }
 
@@ -201,7 +213,8 @@ typedef void (*MuContainerPathForeachFunc) (MuContainer*, gpointer, Path*);
 
 static void
 mu_container_path_foreach_real (MuContainer *c, guint level, Path *path,
-			     MuContainerPathForeachFunc func, gpointer user_data)
+				MuContainerPathForeachFunc func,
+				gpointer user_data)
 {
 	if (!c)
 		return;
@@ -210,7 +223,8 @@ mu_container_path_foreach_real (MuContainer *c, guint level, Path *path,
 	func (c, user_data, path);
 
 	/* children */
-	mu_container_path_foreach_real (c->child, level + 1, path, func, user_data);
+	mu_container_path_foreach_real (c->child, level + 1, path,
+					func, user_data);
 
 	/* siblings */
 	mu_container_path_foreach_real (c->next, level, path, func, user_data);
diff --git a/lib/mu-container.h b/lib/mu-container.h
index b0fc357e..b309462b 100644
--- a/lib/mu-container.h
+++ b/lib/mu-container.h
@@ -29,8 +29,7 @@ enum _MuContainerFlag {
 	MU_CONTAINER_FLAG_SPLICE  = 1 << 1,
 	MU_CONTAINER_FLAG_DUP     = 1 << 2
 };
-
-typedef guint8 MuContainerFlag;
+typedef enum _MuContainerFlag MuContainerFlag;
 
 /*
  * MuContainer data structure, as seen in JWZs document:
@@ -38,10 +37,21 @@ typedef guint8 MuContainerFlag;
  */
 struct _MuContainer {
 	struct _MuContainer *parent, *child, *next;
-	MuContainerFlag flags;
-	MuMsg *msg;
-	guint docid;
-	const char* msgid;
+
+	/* note: we cache the last of the string of next->next->...
+	 * `mu_container_append_siblings' shows up high in the
+	 * profiles since it needs to walk to the end, and this give
+	 * O(n*n) behavior.
+	 * */
+	struct _MuContainer *last;
+
+
+	MuMsg               *msg;
+	const char          *msgid;
+
+	unsigned            docid;
+	MuContainerFlag     flags;
+
 };
 typedef struct _MuContainer MuContainer;
 
@@ -177,7 +187,8 @@ typedef int (*MuContainerCmpFunc) (MuContainer *c1, MuContainer *c2,
  *
  * @return a sorted container
  */
-MuContainer* mu_container_sort (MuContainer *c, MuMsgFieldId mfid, gboolean revert,
+MuContainer* mu_container_sort (MuContainer *c, MuMsgFieldId mfid,
+				gboolean revert,
 				gpointer user_data);