diff --git a/NEWS.md b/NEWS.md
index 5b3ea839..773aa8b8 100644
--- a/NEWS.md
+++ b/NEWS.md
@@ -1,5 +1,6 @@
 # mlr3fselect (development version)
 
+* feat: Features can be always included with the `always_include` column role.
 * fix: Add `$phash()` method to `AutoFSelector`.
 * fix: Include `FSelector` in hash of  `AutoFSelector`.
 * refactor: Change default batch size of `FSelectorRandomSearch` to 10.
diff --git a/R/ObjectiveFSelect.R b/R/ObjectiveFSelect.R
index 79f1b632..52bfabc8 100644
--- a/R/ObjectiveFSelect.R
+++ b/R/ObjectiveFSelect.R
@@ -82,7 +82,9 @@ ObjectiveFSelect = R6Class("ObjectiveFSelect",
       tasks = map(private$.xss, function(x) {
         state = self$task$feature_names[unlist(x)]
         task = self$task$clone()
-        task$select(state)
+        always_included = task$col_roles$always_included
+        task$set_col_roles(always_included, "feature")
+        task$select(c(state, always_included))
         task
       })
 
diff --git a/R/zzz.R b/R/zzz.R
index 1830119a..0ec1c695 100644
--- a/R/zzz.R
+++ b/R/zzz.R
@@ -16,6 +16,10 @@
   x = utils::getFromNamespace("bbotk_reflections", ns = "bbotk")
   x$optimizer_properties = c(x$optimizer_properties, "requires_model")
 
+  x = utils::getFromNamespace("mlr_reflections", ns = "mlr3")
+  x$task_col_roles$classif = c(x$task_col_roles$classif, "always_included")
+  x$task_col_roles$regr = c(x$task_col_roles$regr, "always_included")
+
   # callbacks
   x = utils::getFromNamespace("mlr_callbacks", ns = "mlr3misc")
   x$add("mlr3fselect.backup", load_callback_backup)
diff --git a/tests/testthat/test_FSelectInstanceSingleCrit.R b/tests/testthat/test_FSelectInstanceSingleCrit.R
index 171f291d..548738e5 100644
--- a/tests/testthat/test_FSelectInstanceSingleCrit.R
+++ b/tests/testthat/test_FSelectInstanceSingleCrit.R
@@ -66,3 +66,55 @@ test_that("result$features works", {
   inst$assign_result(xdt, y)
   expect_character(inst$result_feature_set)
 })
+
+test_that("always include variable works", {
+  task = tsk("pima")
+  task$set_col_roles("glucose", "always_included")
+
+  learner = lrn("classif.rpart")
+  resampling = rsmp("cv", folds = 3)
+
+  instance = fselect(
+    fselector = fs("random_search", batch_size = 100),
+    task = task,
+    learner = learner,
+    resampling = resampling,
+    measure = msr("classif.ce"),
+    terminator = trm("evals", n_evals = 100),
+    store_models = TRUE
+  )
+
+  data = as.data.table(instance$archive)
+
+  expect_names(instance$archive$cols_x, disjunct.from = "gloucose")
+  expect_names(names(instance$archive$data), disjunct.from = "gloucose")
+  walk(data$resample_result, function(rr) {
+    expect_names(names(rr$learners[[1]]$state$data_prototype), must.include = "glucose")
+  })
+})
+
+test_that("always include variables works", {
+  task = tsk("pima")
+  task$set_col_roles(c("glucose", "age"), "always_included")
+
+  learner = lrn("classif.rpart")
+  resampling = rsmp("cv", folds = 3)
+
+  instance = fselect(
+    fselector = fs("random_search", batch_size = 100),
+    task = task,
+    learner = learner,
+    resampling = resampling,
+    measure = msr("classif.ce"),
+    terminator = trm("evals", n_evals = 100),
+    store_models = TRUE
+  )
+
+  data = as.data.table(instance$archive)
+
+  expect_names(instance$archive$cols_x, disjunct.from = c("glucose", "age"))
+  expect_names(names(instance$archive$data), disjunct.from = c("glucose", "age"))
+  walk(data$resample_result, function(rr) {
+    expect_names(names(rr$learners[[1]]$state$data_prototype), must.include = c("glucose", "age"))
+  })
+})