diff --git a/src/hackney.erl b/src/hackney.erl
index 2930c560..e1097bd3 100644
--- a/src/hackney.erl
+++ b/src/hackney.erl
@@ -1095,7 +1095,8 @@ async_request(ConnPid, Method, Path, Headers, Body, AsyncMode, StreamTo, FollowR
   %% Handle body encoding
   {FinalHeaders, FinalBody} = encode_body(Headers, Body, []),
   HeadersList = hackney_headers:to_list(FinalHeaders),
-
+  %% Note: Issue #646 - ownership transfer to StreamTo (when different from caller)
+  %% is handled atomically inside hackney_conn:do_request_async
   case hackney_conn:request_async(ConnPid, Method, Path, HeadersList, FinalBody, AsyncMode, StreamTo, FollowRedirect) of
     {ok, Ref} ->
       {ok, Ref};
diff --git a/src/hackney_conn.erl b/src/hackney_conn.erl
index 7c3962cd..9799188a 100644
--- a/src/hackney_conn.erl
+++ b/src/hackney_conn.erl
@@ -1930,7 +1930,20 @@ notify_pool_available_sync(#conn_data{pool_pid = PoolPid, upgraded_ssl = Upgrade
 do_request_async(From, Method, Path, Headers, Body, AsyncMode, StreamTo, FollowRedirect, Data) ->
     %% Use self() (connection PID) as the async ref for message correlation
     Ref = self(),
-    NewData = Data#conn_data{
+    %% Issue #646: If StreamTo is different from the caller, set StreamTo as owner.
+    %% This way the connection lifecycle is tied to the message recipient.
+    %% If StreamTo dies, the connection will be terminated.
+    {CallerPid, _} = From,
+    Data2 = case StreamTo =:= CallerPid of
+        true ->
+            Data;
+        false ->
+            %% Change owner to StreamTo
+            demonitor(Data#conn_data.owner_mon, [flush]),
+            NewMon = monitor(process, StreamTo),
+            Data#conn_data{owner = StreamTo, owner_mon = NewMon}
+    end,
+    NewData = Data2#conn_data{
         request_from = From,
         method = Method,
         path = Path,
diff --git a/test/hackney_async_tests.erl b/test/hackney_async_tests.erl
new file mode 100644
index 00000000..aabb5916
--- /dev/null
+++ b/test/hackney_async_tests.erl
@@ -0,0 +1,217 @@
+%% @doc Tests for async request ownership (issue #646)
+%% When stream_to is provided, the stream_to process should be the owner.
+-module(hackney_async_tests).
+-include_lib("eunit/include/eunit.hrl").
+
+-define(PORT, 9878).
+
+%% Setup/teardown for integration tests
+setup() ->
+    {ok, _} = application:ensure_all_started(hackney),
+    {ok, _} = application:ensure_all_started(cowboy),
+    Dispatch = cowboy_router:compile([{'_', [{"/[...]", test_http_resource, []}]}]),
+    {ok, _} = cowboy:start_clear(test_async_http, [{port, ?PORT}], #{
+        env => #{dispatch => Dispatch}
+    }),
+    ok.
+
+cleanup(_) ->
+    cowboy:stop_listener(test_async_http),
+    ok.
+
+url(Path) ->
+    <<"http://localhost:", (integer_to_binary(?PORT))/binary, Path/binary>>.
+
+%% =============================================================================
+%% Issue #646: stream_to as owner for async requests
+%% =============================================================================
+
+stream_to_owner_test_() ->
+    {setup,
+     fun setup/0,
+     fun cleanup/1,
+     [
+      {"stream_to becomes connection owner", fun test_stream_to_becomes_owner/0},
+      {"connection terminates when stream_to dies", fun test_connection_terminates_when_stream_to_dies/0},
+      {"original caller death does not affect connection when stream_to is different",
+       fun test_caller_death_with_different_stream_to/0},
+      {"backward compat: stream_to = self() works normally", fun test_stream_to_self_compat/0}
+     ]}.
+
+%% Test that when stream_to is provided, it becomes the connection owner
+test_stream_to_becomes_owner() ->
+    %% Create a separate receiver process
+    Self = self(),
+    Receiver = spawn_link(fun() -> receiver_loop(Self) end),
+
+    %% Make async request with stream_to = Receiver
+    {ok, Ref} = hackney:request(get, url(<<"/get">>), [], <<>>,
+                                [{async, true}, {stream_to, Receiver}]),
+    ?assert(is_pid(Ref)),
+
+    %% Wait for receiver to collect all messages
+    timer:sleep(500),
+
+    %% Ask receiver for messages received
+    Receiver ! {get_messages, self()},
+    Messages = receive
+        {messages, Msgs} -> Msgs
+    after 2000 -> []
+    end,
+
+    %% Should have received status and headers
+    ?assert(lists:any(fun({status, S, _}) -> S >= 200; (_) -> false end, Messages)),
+    ?assert(lists:any(fun({headers, _}) -> true; (_) -> false end, Messages)),
+    ?assert(lists:member(done, Messages)),
+
+    %% Clean up
+    Receiver ! stop,
+    ok.
+
+%% Test that connection terminates when stream_to process dies
+test_connection_terminates_when_stream_to_dies() ->
+    %% Create a receiver process that will die on command
+    Self = self(),
+    Receiver = spawn(fun() ->
+        receive
+            die -> exit(killed);
+            {get_messages, Pid} -> Pid ! {messages, []}, receiver_loop(Self)
+        end
+    end),
+
+    %% Make async request with stream_to = Receiver
+    %% Use a large chunked response to ensure we have time to kill the receiver
+    {ok, Ref} = hackney:request(get, url(<<"/chunked/1000000">>), [], <<>>,
+                                [{async, true}, {stream_to, Receiver}]),
+    ?assert(is_pid(Ref)),
+
+    %% Give time for ownership transfer to complete
+    timer:sleep(200),
+
+    %% Connection should be alive
+    ?assert(is_process_alive(Ref)),
+
+    %% Kill the receiver (stream_to)
+    Receiver ! die,
+    timer:sleep(200),
+    ?assertNot(is_process_alive(Receiver)),
+
+    %% Connection should be terminated because stream_to died
+    %% Wait a bit longer for the monitor DOWN message to propagate
+    timer:sleep(500),
+    ?assertNot(is_process_alive(Ref)),
+    ok.
+
+%% Test that when stream_to is different from caller, caller death doesn't affect connection
+test_caller_death_with_different_stream_to() ->
+    %% We spawn a process to make the request, then kill it
+    %% The receiver should still get all messages
+    Self = self(),
+    Receiver = spawn_link(fun() -> receiver_loop(Self) end),
+
+    %% Spawn a caller that makes the request then dies
+    Caller = spawn(fun() ->
+        {ok, Ref} = hackney:request(get, url(<<"/get">>), [], <<>>,
+                                    [{async, true}, {stream_to, Receiver}]),
+        Self ! {started, Ref},
+        %% Die immediately after starting request
+        exit(normal)
+    end),
+
+    %% Wait for caller to start the request and get the ref
+    ConnRef = receive
+        {started, R} -> R
+    after 2000 -> error(timeout_waiting_for_start)
+    end,
+
+    %% Wait for caller to die
+    timer:sleep(100),
+    ?assertNot(is_process_alive(Caller)),
+
+    %% Connection should still be alive (owned by Receiver now)
+    ?assert(is_process_alive(ConnRef)),
+
+    %% Give it some time to complete the request
+    timer:sleep(500),
+
+    %% Ask receiver for messages
+    Receiver ! {get_messages, self()},
+    Messages = receive
+        {messages, Msgs} -> Msgs
+    after 2000 -> []
+    end,
+
+    %% Should have received complete response despite caller death
+    ?assert(lists:any(fun({status, S, _}) -> S >= 200; (_) -> false end, Messages)),
+    ?assert(lists:any(fun({headers, _}) -> true; (_) -> false end, Messages)),
+    ?assert(lists:member(done, Messages)),
+
+    %% Clean up
+    Receiver ! stop,
+    ok.
+
+%% Test backward compatibility: when stream_to = self(), behavior unchanged
+test_stream_to_self_compat() ->
+    %% When stream_to is self (default), messages come to caller
+    {ok, Ref} = hackney:request(get, url(<<"/get">>), [], <<>>,
+                                [{async, true}]),  %% Default stream_to = self()
+    ?assert(is_pid(Ref)),
+
+    %% Should receive status
+    receive
+        {hackney_response, Ref, {status, Status, _Reason}} ->
+            ?assert(Status >= 200 andalso Status < 400)
+    after 5000 ->
+        ?assert(false)
+    end,
+
+    %% Should receive headers
+    receive
+        {hackney_response, Ref, {headers, Headers}} ->
+            ?assert(is_list(Headers))
+    after 5000 ->
+        ?assert(false)
+    end,
+
+    %% Collect remaining messages
+    Messages = receive_all_async(Ref, []),
+    ?assert(lists:member(done, Messages)),
+
+    ok.
+
+%% =============================================================================
+%% Helper functions
+%% =============================================================================
+
+%% A receiver process that collects async messages
+receiver_loop(Parent) ->
+    receiver_loop(Parent, []).
+
+receiver_loop(Parent, Acc) ->
+    receive
+        {hackney_response, _Ref, done} ->
+            receiver_loop(Parent, [done | Acc]);
+        {hackney_response, _Ref, {error, _} = Error} ->
+            receiver_loop(Parent, [Error | Acc]);
+        {hackney_response, _Ref, Msg} ->
+            receiver_loop(Parent, [Msg | Acc]);
+        {get_messages, Pid} ->
+            Pid ! {messages, lists:reverse(Acc)},
+            receiver_loop(Parent, Acc);
+        stop ->
+            ok
+    after 30000 ->
+        ok
+    end.
+
+receive_all_async(Ref, Acc) ->
+    receive
+        {hackney_response, Ref, done} ->
+            lists:reverse([done | Acc]);
+        {hackney_response, Ref, {error, _} = Error} ->
+            lists:reverse([Error | Acc]);
+        {hackney_response, Ref, Msg} ->
+            receive_all_async(Ref, [Msg | Acc])
+    after 5000 ->
+        lists:reverse(Acc)
+    end.