diff --git a/lib/pls/include/pls/internal/data_structures/delayed_initialization_wrapper.h b/lib/pls/include/pls/internal/data_structures/delayed_initialization_wrapper.h
index 1d93398..d02f94e 100644
--- a/lib/pls/include/pls/internal/data_structures/delayed_initialization_wrapper.h
+++ b/lib/pls/include/pls/internal/data_structures/delayed_initialization_wrapper.h
@@ -35,11 +35,19 @@ class delayed_initialization_wrapper {
 
   template<typename ...ARGS>
   void initialize(ARGS &&...args) {
-    PLS_ASSERT(initialized_, "Can only initialize delayed wrapper object once!")
+    PLS_ASSERT(!initialized_, "Can only initialize delayed wrapper object once!")
 
     new(memory_) T(std::forward<ARGS>(args)...);
     initialized_ = true;
   }
+
+  void destroy() {
+    PLS_ASSERT(initialized_, "Can only destroy initialized objects!")
+
+    memory_->~T();
+    initialized_ = false;
+  }
+
   T &object() {
     PLS_ASSERT(initialized_, "Can not use an uninitialized delayed wrapper object!")
     return *reinterpret_cast<T *>(memory_);
diff --git a/lib/pls/include/pls/internal/scheduling/cont_manager.h b/lib/pls/include/pls/internal/scheduling/cont_manager.h
index 104337f..1290921 100644
--- a/lib/pls/include/pls/internal/scheduling/cont_manager.h
+++ b/lib/pls/include/pls/internal/scheduling/cont_manager.h
@@ -3,6 +3,8 @@
 #define PLS_CONT_MANAGER_H_
 
 #include <memory>
+#include <tuple>
+#include <array>
 
 #include "pls/internal/data_structures/aligned_stack.h"
 #include "pls/internal/scheduling/continuation.h"
@@ -13,14 +15,45 @@ namespace scheduling {
 
 class cont_manager {
  public:
-  explicit cont_manager(size_t num_conts, size_t max_cont_size, data_structures::aligned_stack &cont_storage)
-      : num_conts_{num_conts}, max_cont_size_{max_cont_size}, cont_storage_{cont_storage} {
-    //TODO: Init linked list like structure
+  template<size_t NUM_CONTS, size_t MAX_CONT_SIZE>
+  explicit cont_manager(data_structures::aligned_stack &cont_storage) {
+    // First node is currently active and our local start
+    start_node_ = active_node_ = init_cont_node<MAX_CONT_SIZE>(cont_storage, nullptr, nullptr);
+
+    // Build up chain after it
+    continuation_node *current_node = start_node_;
+    for (size_t i = 1; i < NUM_CONTS; i++) {
+      continuation_node *next_node = init_cont_node<MAX_CONT_SIZE>(cont_storage, start_node_, current_node);
+      current_node->set_prev(next_node);
+    }
   };
 
+  continuation_node *fast_path_get_next() {
+    active_node_ = active_node_->get_next();
+    return active_node_;
+  }
+  void fast_path_return() {
+    active_node_ = active_node_->get_prev();
+  }
+
+ private:
+  template<size_t MAX_CONT_SIZE>
+  static continuation_node *init_cont_node(data_structures::aligned_stack &cont_storage,
+                                           continuation_node *cont_chain_start,
+                                           continuation_node *prev) {
+    // Represents one cont node and its corresponding memory buffer (as one continuous block of memory).
+    using cont_node_memory_pair = std::tuple<continuation_node,
+                                             std::array<char, MAX_CONT_SIZE - sizeof(continuation_node)>>;
+    char *tuple_memory = cont_storage.push_bytes<cont_node_memory_pair>();
+    char *cont_node_address = tuple_memory;
+    char *cont_node_memory_address = tuple_memory + sizeof(continuation_node);
+
+    return new(cont_node_address) continuation_node(cont_node_memory_address, cont_chain_start, prev);
+  }
+
  private:
-  const size_t num_conts_, max_cont_size_;
-  data_structures::aligned_stack &cont_storage_;
+  continuation_node *start_node_;
+  continuation_node *active_node_;
 };
 
 template<size_t NUM_CONTS, size_t MAX_CONT_SIZE>
diff --git a/lib/pls/include/pls/internal/scheduling/continuation.h b/lib/pls/include/pls/internal/scheduling/continuation.h
index f9a2420..6a2ef5f 100644
--- a/lib/pls/include/pls/internal/scheduling/continuation.h
+++ b/lib/pls/include/pls/internal/scheduling/continuation.h
@@ -20,10 +20,27 @@ class base_continuation {
 
 class continuation_node {
  public:
+  continuation_node(char *continuation_memory, continuation_node *cont_chain_start, continuation_node *prev)
+      : continuation_memory_{continuation_memory},
+        continuation_{nullptr},
+        cont_chain_start_{cont_chain_start},
+        prev_{prev},
+        next_{nullptr},
+        offered_chain_{nullptr} {}
+
+  void set_cont_chain_start(continuation_node *cont_chain_start) { cont_chain_start_ = cont_chain_start; }
+  continuation_node *get_cont_chain_start() { return cont_chain_start_; }
+  void set_next(continuation_node *next) { next_ = next; }
+  continuation_node *get_next() { return next_; }
+  void set_prev(continuation_node *prev) { prev_ = prev; }
+  continuation_node *get_prev() { return prev_; }
+
+  base_continuation *continuation() { return continuation_; }
 
  private:
   // Pointer to memory region reserved for the companion continuation.
   // Must be a buffer big enough to hold any continuation encountered in the program.
+  char *continuation_memory_;
   base_continuation *continuation_;
 
   // Linked list property of continuations (continuation chains as memory management).
@@ -47,6 +64,10 @@ class continuation : public base_continuation {
     //       E.g. handle passing the result to the parent continuation
     function_.object()(result_1_.object(), result_2_.object());
   }
+
+  // Warning: De-constructor only called once the continuation is actually needed.
+  //          This should be a non issue, as all memory in the continuation is uninitialized
+  //          until the first use anyways to save runtime.
   ~continuation() override = default;
 
   template<typename R1ARG>
@@ -76,14 +97,15 @@ class continuation : public base_continuation {
   template<typename T>
   using delayed_init = data_structures::delayed_initialization_wrapper<T>;
 
-  delayed_init<R1> result_1_;
-  delayed_init<R2> result_2_;
-  delayed_init<F> function_;
-
   // Also uninitialized at first, only take the atomic write on the slow path.
   // The stealer will init it to 2 while stealing, the 'stolen' sync will then make sure
   // everyone sees the value in correct order.
   std::atomic<unsigned short> results_missing_{};
+
+  // All fields/actual values stay uninitialized (save time on the fast path if we don not need them).
+  delayed_init<R1> result_1_;
+  delayed_init<R2> result_2_;
+  delayed_init<F> function_;
 };
 
 }