diff --git a/README.md b/README.md
index 28475db..ee6a019 100644
--- a/README.md
+++ b/README.md
@@ -89,46 +89,5 @@ assert(Immutable{a=1,b=2,c=3} == Immutable{a=1,b=2,c=3})
 This library is pretty dang fast, but it's still slower than native Lua tables. Based on my local testing, immutable tables add a couple nanoseconds to table operations in the worst case scenario. In LuaJIT, there is a bigger performance discrepancy because regular tables are more heavily optimized in LuaJIT. Your mileage may vary, but I'd say that immutable tables will probably never be a performance bottleneck for your program, especially if you use them in place of code that already used a constructor function and metatables. In some cases, immutable tables may also help reduce your program's memory footprint (if your program has many duplicate objects in memory) and may even improve speed (e.g. if your program uses a lot of deep equality checks). Don't trust this paragraph though! If in doubt, profile your code!
 
 ## Implementation details
-Under the hood, immutable tables are implemented in C as a userdata (that stores a hash value and length) with a metatable. That metatable has a weak-keyed mapping from hash -> userdata -> associated data. Immutable tables *are* garbage collected, so if you no longer have any references to the userdata, the userdata will get garbage collected, which will result in the entry being removed from the metatable's mapping. When new instances are created, a hash value is computed, and all the data in the associated hash bucket is scanned to look for duplicates. If a match is found, the existing instance is returned, otherwise a new one is created and added to the hash bucket. The following lua pseudocode approximates the C implementation's behavior:
-```lua
-function immutable(fields, class_fields)
-    local cls = {
-        __index = function(self, key)
-            local cls = getmetatable(self)
-            key = cls.__indices[key] or key
-            local data = cls.__instances[extract_hash(self)][self]
-            return data[key] or cls[key]
-        end,
-        ...also: __len, __pairs, __ipairs, __tostring, from_table, is_instance...
-    }
-    for k,v in pairs(class_fields) do cls[k] = v end
-    cls.__fields, cls.__indices = fields, {}
-    for i,f in ipairs(fields) do cls.__indices[f] = i end
-    cls.__instances = setmetatable({}, {__mode='v', __index=function(self, key)
-        local bucket = setmetatable({}, {__mode='k'})
-        self[key] = bucket
-        return bucket
-    end})
-    cls.__buckets = setmetatable({}, {__mode='k'})
-    return setmetatable(cls, {
-        __call=function(cls, ...)
-            local hash = calculate_hash(...)
-            local bucket = cls.__instances[hash]
-            for userdata, data in pairs(bucket) do
-                local match = true
-                for i=1,select('#',...) do
-                    if data[i] ~= select(i, ...) then
-                        match = false
-                        break
-                    end
-                end
-                if match then return userdata end
-            end
-            local userdata = setmetatable(new_userdata(hash), cls)
-            cls.__buckets[userdata] = bucket
-            bucket[userdata] = {...}
-            return userdata
-        end
-    })
-end
-```
+
+You can read more about the implementation details in the [implementation documentation](./implementation.md).
diff --git a/implementation.md b/implementation.md
new file mode 100644
index 0000000..e3495d3
--- /dev/null
+++ b/implementation.md
@@ -0,0 +1,44 @@
+# Implementation details
+Under the hood, immutable tables are implemented in C as a userdata (that stores a hash value and length) with a metatable. That metatable has a weak-keyed mapping from hash -> userdata -> associated data. Immutable tables *are* garbage collected, so if you no longer have any references to the userdata, the userdata will get garbage collected, which will result in the entry being removed from the metatable's mapping. When new instances are created, a hash value is computed, and all the data in the associated hash bucket is scanned to look for duplicates. If a match is found, the existing instance is returned, otherwise a new one is created and added to the hash bucket. The following lua pseudocode approximates the C implementation's behavior:
+```lua
+function immutable(fields, class_fields)
+    local cls = {
+        __index = function(self, key)
+            local cls = getmetatable(self)
+            key = cls.__indices[key] or key
+            local data = cls.__instances[extract_hash(self)][self]
+            return data[key] or cls[key]
+        end,
+        ...also: __len, __pairs, __ipairs, __tostring, from_table, is_instance...
+    }
+    for k,v in pairs(class_fields) do cls[k] = v end
+    cls.__fields, cls.__indices = fields, {}
+    for i,f in ipairs(fields) do cls.__indices[f] = i end
+    cls.__instances = setmetatable({}, {__mode='v', __index=function(self, key)
+        local bucket = setmetatable({}, {__mode='k'})
+        self[key] = bucket
+        return bucket
+    end})
+    cls.__buckets = setmetatable({}, {__mode='k'})
+    return setmetatable(cls, {
+        __call=function(cls, ...)
+            local hash = calculate_hash(...)
+            local bucket = cls.__instances[hash]
+            for userdata, data in pairs(bucket) do
+                local match = true
+                for i=1,select('#',...) do
+                    if data[i] ~= select(i, ...) then
+                        match = false
+                        break
+                    end
+                end
+                if match then return userdata end
+            end
+            local userdata = setmetatable(new_userdata(hash), cls)
+            cls.__buckets[userdata] = bucket
+            bucket[userdata] = {...}
+            return userdata
+        end
+    })
+end
+```