diff --git a/lib/pls/include/pls/algorithms/scan_impl.h b/lib/pls/include/pls/algorithms/scan_impl.h
index 5abe558..1af0418 100644
--- a/lib/pls/include/pls/algorithms/scan_impl.h
+++ b/lib/pls/include/pls/algorithms/scan_impl.h
@@ -93,6 +93,9 @@ class scan_task : public pls::internal::scheduling::task {
         }
       }
     }, fixed_strategy{1});
+
+    wait_for_all();
+    this->~scan_task();
   }
 };
 
diff --git a/lib/pls/include/pls/dataflow/internal/function_node.h b/lib/pls/include/pls/dataflow/internal/function_node.h
index d7490b7..47fed95 100644
--- a/lib/pls/include/pls/dataflow/internal/function_node.h
+++ b/lib/pls/include/pls/dataflow/internal/function_node.h
@@ -105,6 +105,10 @@ class function_node<inputs<I0, I...>, outputs<O0, O...>, F> : public node {
     auto invocation = new(memory) invocation_memory{};
     invocation->inputs_missing_ = num_in_ports;
   };
+  void clean_up_instance_buffer(void *memory) const override {
+    auto invocation = reinterpret_cast<invocation_memory *>(memory);
+    invocation->~invocation_memory();
+  }
 
  private:
   multi_in_port_type in_port_;
diff --git a/lib/pls/include/pls/dataflow/internal/graph.h b/lib/pls/include/pls/dataflow/internal/graph.h
index fa984dc..d1580b8 100644
--- a/lib/pls/include/pls/dataflow/internal/graph.h
+++ b/lib/pls/include/pls/dataflow/internal/graph.h
@@ -115,6 +115,10 @@ class graph<inputs<I0, I...>, outputs<O0, O...>> : public node {
   void init_instance_buffer(void *memory) const override {
     new(memory) invocation_memory{};
   };
+  void clean_up_instance_buffer(void *memory) const override {
+    auto invocation = reinterpret_cast<invocation_memory *>(memory);
+    invocation->~invocation_memory();
+  }
 
  private:
   inputs_type inputs_;
diff --git a/lib/pls/include/pls/dataflow/internal/graph_impl.h b/lib/pls/include/pls/dataflow/internal/graph_impl.h
index 377673f..f135982 100644
--- a/lib/pls/include/pls/dataflow/internal/graph_impl.h
+++ b/lib/pls/include/pls/dataflow/internal/graph_impl.h
@@ -129,9 +129,22 @@ class graph<inputs<I0, I...>, outputs<O0, O...>>::run_graph_task : public pls::t
     self_->get_invocation<invocation_memory>(invocation_)->output_buffer_ = output;
   }
 
+  ~run_graph_task() {
+    node *iterator = self_->node_list_start_;
+    for (int i = 0; i < self_->num_nodes_; i++) {
+      void* memory = invocation_.get_instance_buffer<void>(i);
+      iterator->clean_up_instance_buffer(memory);
+
+      iterator = iterator->direct_successor_;
+    }
+  }
+
   void execute_internal() override {
     PROFILE_WORK_BLOCK("Graph Invocation")
     feed_inputs<0, I0, I...>{self_->inputs_, input_, invocation_}.run();
+    
+    wait_for_all();
+    this->~run_graph_task();
     PROFILE_END_BLOCK
   }
 };
diff --git a/lib/pls/include/pls/dataflow/internal/merge_node.h b/lib/pls/include/pls/dataflow/internal/merge_node.h
index 2a7a6c5..8da2e64 100644
--- a/lib/pls/include/pls/dataflow/internal/merge_node.h
+++ b/lib/pls/include/pls/dataflow/internal/merge_node.h
@@ -102,6 +102,10 @@ class merge_node : public node {
     auto invocation = new(memory) invocation_memory{};
     invocation->inputs_missing_ = INITIAL_STATE;
   };
+  void clean_up_instance_buffer(void *memory) const override {
+    auto invocation = reinterpret_cast<invocation_memory *>(memory);
+    invocation->~invocation_memory();
+  }
 
  private:
   multi_in_port_type in_port_;
diff --git a/lib/pls/include/pls/dataflow/internal/node.h b/lib/pls/include/pls/dataflow/internal/node.h
index 0b9c447..037a3eb 100644
--- a/lib/pls/include/pls/dataflow/internal/node.h
+++ b/lib/pls/include/pls/dataflow/internal/node.h
@@ -25,6 +25,7 @@ class node {
 
   virtual int instance_buffer_size() const = 0;
   virtual void init_instance_buffer(void *memory) const = 0;
+  virtual void clean_up_instance_buffer(void *memory) const = 0;
 
   virtual bool is_fully_connected() const = 0;
 
diff --git a/lib/pls/include/pls/dataflow/internal/split_node.h b/lib/pls/include/pls/dataflow/internal/split_node.h
index c0956d2..363f32c 100644
--- a/lib/pls/include/pls/dataflow/internal/split_node.h
+++ b/lib/pls/include/pls/dataflow/internal/split_node.h
@@ -60,6 +60,9 @@ class split_node : public node {
   void init_instance_buffer(void *) const override {
     // No need for memory, we simply forward entries without buffering
   };
+  void clean_up_instance_buffer(void *memory) const override {
+    // No need for memory, we simply forward entries without buffering
+  }
 
  private:
   multi_in_port_type in_port_;
diff --git a/lib/pls/include/pls/dataflow/internal/switch_node.h b/lib/pls/include/pls/dataflow/internal/switch_node.h
index 598275b..b371c0b 100644
--- a/lib/pls/include/pls/dataflow/internal/switch_node.h
+++ b/lib/pls/include/pls/dataflow/internal/switch_node.h
@@ -86,6 +86,10 @@ class switch_node : public node {
     auto invocation = new(memory) invocation_memory{};
     invocation->inputs_missing_ = 2;
   };
+  void clean_up_instance_buffer(void *memory) const override {
+    auto invocation = reinterpret_cast<invocation_memory *>(memory);
+    invocation->~invocation_memory();
+  }
 
  private:
   multi_in_port_type in_port_;
diff --git a/lib/pls/include/pls/internal/scheduling/lambda_task.h b/lib/pls/include/pls/internal/scheduling/lambda_task.h
index fb6cc4a..1d3d4b1 100644
--- a/lib/pls/include/pls/internal/scheduling/lambda_task.h
+++ b/lib/pls/include/pls/internal/scheduling/lambda_task.h
@@ -18,6 +18,9 @@ class lambda_task_by_reference : public task {
  protected:
   void execute_internal() override {
     function_();
+
+    wait_for_all();
+    this->~lambda_task_by_reference<Function>();
   }
 };
 
@@ -31,6 +34,9 @@ class lambda_task_by_value : public task {
  protected:
   void execute_internal() override {
     function_();
+
+    wait_for_all();
+    this->~lambda_task_by_value<Function>();
   }
 };
 
diff --git a/lib/pls/include/pls/internal/scheduling/task.h b/lib/pls/include/pls/internal/scheduling/task.h
index e8925cb..e7b23fb 100644
--- a/lib/pls/include/pls/internal/scheduling/task.h
+++ b/lib/pls/include/pls/internal/scheduling/task.h
@@ -13,6 +13,18 @@ namespace pls {
 namespace internal {
 namespace scheduling {
 
+/**
+ * A task to be executed by the runtime system.
+ * Tasks are guaranteed to be executed exactly once.
+ *
+ * Override the execute_internal() method for your custom code.
+ *
+ * IMPORTANT:
+ * Tasks memory is re-used without calling the destructor.
+ * You must call it yourself at the end of execute_internal().
+ * This is done to not introduce any overhead of virtual function calls
+ * if no clean up is required.
+ */
 class task {
   friend class scheduler;
 
@@ -29,6 +41,10 @@ class task {
  protected:
   /*
    * Must call the parent constructor.
+   *
+   * IMPORTANT:
+   * Tasks memory is re-used without calling the destructor.
+   * You must call it yourself at the end of execute_internal().
    */
   explicit task();
 
@@ -37,6 +53,9 @@ class task {
    * Memory will be pushed onto the stack (in aligned memory, thus avoid many small chunks).
    * MUST be called in constructor, never afterwards.
    *
+   * Memory is fully self managed. Calling e.g. deconstructors when not needing objects
+   * anymore is the users responsibility (memory is simply re-used after the life time of the task ends).
+   *
    * @param size Number of bytes to be allocated
    * @return The allocated memory region
    */