diff --git a/libs/arr_ref/arr_ref.nim b/libs/arr_ref/arr_ref.nim
index 3bb9d88..1a0be3d 100644
--- a/libs/arr_ref/arr_ref.nim
+++ b/libs/arr_ref/arr_ref.nim
@@ -1,12 +1,5 @@
 import std/strformat
 import std/hashes
-import ./slice_mem
-export slice_mem
-
-const header_size = when not defined(release):
-    sizeof(pointer)*2 + sizeof(Natural)
-else:
-    sizeof(pointer)
 
 type ArrRef*[T] = ref object
     byte_len: int
@@ -15,6 +8,8 @@ type ArrRef*[T] = ref object
         arr_ptr: ptr T
     arr: UncheckedArray[T]
 
+const header_size =  offsetof(ArrRef[byte], arr)
+
 proc newArrRef*[T](size: Natural): ArrRef[T] =
     when defined(TCC):
         var r: ref pointer # workaround for tcc bug
@@ -130,9 +125,6 @@ proc hash*[T](arr: ArrRef[T]): Hash =
     hash(cast[ptr UncheckedArray[byte]](arr.arr.addr).toOpenArray(0, arr.len * sizeof(T) - 1))
     # TODO: for bigger elements, would a different algorithm be faster?
 
-template `[]`*[T](a: ArrRef[T], s: Slice[int]): var SliceMem[T] =
-    toSliceMem(a, s)
-
 template concatImpl(arrs: untyped) =
     var total = 0
     for a in arrs:
@@ -152,3 +144,9 @@ proc concat*[T](arrs: seq[ArrRef[T]]): ArrRef[T] =
 template `&`*[T,U](a: ArrRef[T], b: ArrRef[U]): ArrRef[T] =
     concat(a, b.to(T))
 
+import ./slice_mem
+export slice_mem
+
+template `[]`*[T](a: ArrRef[T], s: Slice[int]): var SliceMem[T] =
+    toSliceMem(a, s)
+
diff --git a/libs/arr_ref/slice_mem.nim b/libs/arr_ref/slice_mem.nim
index e62d4dc..8835a6d 100644
--- a/libs/arr_ref/slice_mem.nim
+++ b/libs/arr_ref/slice_mem.nim
@@ -7,23 +7,32 @@
 ## 
 ## It's called "SliceMem" because "MemSlice" is taken by std/memfiles.
 
+const arrRefCopyContainers {.booldefine.} = true
+
 import std/strformat
 import std/hashes
 import std/macros # for noMove()
+import ./arr_ref
 
-type
-    SliceMem*[T] = object
+when arrRefCopyContainers:
+    type SliceMem*[T] = object
         data*: ptr UncheckedArray[T]
         byte_len*: int
-        destroy_ref: ref CustomDestructor
-    
-    CustomDestructor = object
-        destroy: proc() {.closure, raises: [].}
+        data_ref: ArrRef[T]
+else:
+    type
+        SliceMem*[T] = object
+            data*: ptr UncheckedArray[T]
+            byte_len*: int
+            destroy_ref: ref CustomDestructor
+        
+        CustomDestructor = object
+            destroy: proc() {.closure, raises: [].}
 
-proc `=destroy`(s: var CustomDestructor) =
-    if s.destroy != nil:
-        s.destroy()
-        s.destroy = nil
+    proc `=destroy`(s: var CustomDestructor) =
+        if s.destroy != nil:
+            s.destroy()
+            s.destroy = nil
 
 template toInt(p: pointer): int = cast[int](p)
 # template toPointer(i: int): pointer = cast[pointer](p)
@@ -31,23 +40,31 @@ template toInt(p: pointer): int = cast[int](p)
 template toPointer*(s: SliceMem): pointer =
     s.data.pointer
 
-proc newSliceMem*[T, U](container: sink U; p: pointer, byte_len: int): SliceMem[T] =
+proc newSliceMem*[T, U](container: sink U; p: pointer, byte_len: Natural): SliceMem[T] =
     ## Create a SliceMem from a container, a pointer, and a length in bytes.
     runnableExamples:
         let x = @[1,2,3,4,5]
         # You can omit the [int] here because the container is iterable
         let s = newSliceMem[int](x, x[0].addr, x.len * sizeof(x[0]))
-    
-    result = SliceMem[T](
-        data: cast[ptr UncheckedArray[T]](p),
-        byte_len: byte_len,
-        destroy_ref: (ref CustomDestructor)(destroy: proc()=
-            # we don't actually need to destroy the container here, destroying
-            # the closure will do it for us. Calling it allows us to use custom
-            # destructors though.
-            discard container
-        ),
-    )
+
+    when arrRefCopyContainers:
+        result = SliceMem[T](
+            byte_len: byte_len,
+            data_ref: newArrRef[T](byte_len div sizeof(T)),
+        )
+        copyMem(result.data_ref.toPointer, p, byte_len)
+        result.data = cast[ptr UncheckedArray[T]](result.data_ref.toPointer)
+    else:
+        result = SliceMem[T](
+            data: cast[ptr UncheckedArray[T]](p),
+            byte_len: byte_len,
+            destroy_ref: (ref CustomDestructor)(destroy: proc()=
+                # we don't actually need to destroy the container here, destroying
+                # the closure will do it for us. Calling it allows us to use custom
+                # destructors though.
+                discard container
+            ),
+        )
 
 proc newSliceMem*[T](p: ptr T, len: int, destructor: proc() {.closure, raises: [].}): SliceMem[T] =
     # Create a SliceMem from a pointer to a type, a length, and a destructor
@@ -57,11 +74,20 @@ proc newSliceMem*[T](p: ptr T, len: int, destructor: proc() {.closure, raises: [
         proc destroy() = deallocShared(x)
         let s = newSliceMem(x, 5, destroy)
     
-    result = SliceMem[T](
-        data: cast[ptr UncheckedArray[T]](p),
-        byte_len: len * sizeof(T),
-        destroy_ref: (ref CustomDestructor)(destroy: destructor),
-    )
+    when arrRefCopyContainers:
+        result = SliceMem[T](
+            byte_len: len * sizeof(T),
+            data_ref: newArrRef[T](len),
+        )
+        copyMem(result.data_ref.toPointer, p, len * sizeof(T))
+        result.data = cast[ptr UncheckedArray[T]](result.data_ref.toPointer)
+        destructor()
+    else:
+        result = SliceMem[T](
+            data: cast[ptr UncheckedArray[T]](p),
+            byte_len: len * sizeof(T),
+            destroy_ref: (ref CustomDestructor)(destroy: destructor),
+        )
 
 proc newSliceMem*(p: pointer, byte_len: int, destructor: proc() {.closure, raises: [].}): SliceMem[byte] {.inline.} =
     ## Create a SliceMem from a pointer without type, a length in bytes, and a
@@ -76,13 +102,22 @@ proc newSliceMem*[T, U](container: sink U; first, last: pointer): SliceMem[T] =
     ## Create a SliceMem from a container and the two pointers of the first and
     ## last elements in the container. Usually you will want to use `toSliceMem`
     ## instead.
-    result = SliceMem[T](
-        data: cast[ptr UncheckedArray[T]](first),
-        byte_len: (last.toInt -% first.toInt) + sizeof(T),
-        destroy_ref: (ref CustomDestructor)(destroy: proc()=
-            discard container
-        ),
-    )
+    let byte_len = (last.toInt -% first.toInt) + sizeof(T)
+    when arrRefCopyContainers:
+        result = SliceMem[T](
+            byte_len: byte_len,
+            data_ref: newArrRef[T](byte_len div sizeof(T)),
+        )
+        copyMem(result.data_ref.toPointer, first, byte_len)
+        result.data = cast[ptr UncheckedArray[T]](result.data_ref.toPointer)
+    else:
+        result = SliceMem[T](
+            data: cast[ptr UncheckedArray[T]](first),
+            byte_len: byte_len,
+            destroy_ref: (ref CustomDestructor)(destroy: proc()=
+                discard container
+            ),
+        )
 
 proc newSliceMem*[T](size: int): SliceMem[T] =
     ## Create a SliceMem from new memory, similar to ArrRef[T].
@@ -91,15 +126,21 @@ proc newSliceMem*[T](size: int): SliceMem[T] =
         for i,n in s.mpairs:
             n = (i+1)*11
         doAssert $s == "SliceMem([11, 22, 33, 44, 55, 66, 77, 88, 99, 110])"
-    var r: ref byte
-    unsafeNew(r, size * sizeof(T))
-    result = SliceMem[T](
-        data: cast[ptr UncheckedArray[T]](r),
-        byte_len: size * sizeof(T),
-        destroy_ref: (ref CustomDestructor)(destroy: proc()=
-            discard r
-        ),
-    )
+    when arrRefCopyContainers:
+        result = SliceMem[T](
+            byte_len: size * sizeof(T),
+            data_ref: newArrRef[T](size),
+        )
+        result.data = cast[ptr UncheckedArray[T]](result.data_ref.toPointer)
+    else:
+        var r = allocShared0 size * sizeof(T)
+        result = SliceMem[T](
+            data: cast[ptr UncheckedArray[T]](r),
+            byte_len: size * sizeof(T),
+            destroy_ref: (ref CustomDestructor)(destroy: proc()=
+                deallocShared r
+            ),
+        )
 
 macro noMove(e: untyped): untyped =
     # remove the "move" from an expression