diff --git a/lib/ecto/changeset.ex b/lib/ecto/changeset.ex
index 0937c4f12d..24c7ca0894 100644
--- a/lib/ecto/changeset.ex
+++ b/lib/ecto/changeset.ex
@@ -289,26 +289,29 @@ defmodule Ecto.Changeset do
         |> Ecto.Changeset.validate_length(...)
 
   Besides the basic types which are mentioned above, such as `:boolean` and `:string`,
-  parameterized types can also be used in schemaless changesets. They implement
-  the `Ecto.ParameterizedType` behaviour and we can create the necessary type info by
-  calling the `init/2` function.
-
-  For example, to use `Ecto.Enum` in a schemaless changeset:
+  embeds and parameterized types can also be used in schemaless changesets.
+  For parameterized types, you can call `Ecto.ParameterizedType.init/2`. For embeds,
+  call `Ecto.Embedded.one/2` or `Ecto.Embedded.many/2` accordingly. Here is an example:
 
       types = %{
         name: :string,
-        role: Ecto.ParameterizedType.init(Ecto.Enum, values: [:reader, :editor, :admin])
+        role: Ecto.ParameterizedType.init(Ecto.Enum, values: [:reader, :editor, :admin]),
+        profile: Ecto.Embedded.one(Profile)
       }
 
-      data  = %{}
-      params = %{name: "Callum", role: "reader"}
+      data = %{profile: nil}
+      params = %{name: "Callum", role: "reader", profile: %{bio: "Hello!"}}
 
       changeset =
         {data, types}
-        |> Ecto.Changeset.cast(params, Map.keys(types))
+        |> Ecto.Changeset.cast(params, [:name, :role])
+        |> Ecto.Changeset.cast_embed(:profile)
         |> Ecto.Changeset.validate_required(...)
         |> Ecto.Changeset.validate_length(...)
 
+  In the example above, `Profile` is an schema that defines a `changeset/2` function.
+  You may instead pass a custom function to the `:with` option of `cast_embed/3`.
+
   Schemaless changesets make Ecto extremely useful to cast, validate and prune data even
   if it is not meant to be persisted to the database.
 
@@ -1387,6 +1390,7 @@ defmodule Ecto.Changeset do
       end
 
     relation = relation!(:cast, type, key, Map.get(types, key))
+    relation = put_relation_context(relation, data, key)
     on_cast = Keyword.get_lazy(opts, :with, fn -> Relation.on_cast_default(relation) end)
     sort = opts_key_from_params(:sort_param, opts, params)
     drop = opts_key_from_params(:drop_param, opts, params)
@@ -1418,9 +1422,7 @@ defmodule Ecto.Changeset do
         missing_relation(changeset, key, required?, relation, opts)
       end
 
-    update_in(changeset.types[key], fn {type, relation} ->
-      {type, %{relation | on_cast: on_cast}}
-    end)
+    put_in(changeset.types[key], {type, %{relation | on_cast: on_cast}})
   end
 
   defp cast_params(:many, nil, sort, drop) when is_list(sort) or is_list(drop) do
@@ -1517,6 +1519,13 @@ defmodule Ecto.Changeset do
           "expected `#{name}` to be an #{type} in `#{op}_#{type}`, got: `#{inspect(schema_type)}`"
   end
 
+  defp put_relation_context(%{field: nil, owner: nil} = relation, data, key) do
+    owner = if is_struct(data), do: data.__struct__, else: nil
+    %{relation | field: key, owner: owner}
+  end
+
+  defp put_relation_context(relation, _data, _key), do: relation
+
   defp force_update(changeset, opts) do
     if Keyword.get(opts, :force_update_on_change, true) do
       put_in(changeset.repo_opts[:force], true)
diff --git a/lib/ecto/embedded.ex b/lib/ecto/embedded.ex
index e8c5fb28fb..cefb204a0b 100644
--- a/lib/ecto/embedded.ex
+++ b/lib/ecto/embedded.ex
@@ -1,6 +1,15 @@
 defmodule Ecto.Embedded do
   @moduledoc """
-  The embedding struct for `embeds_one` and `embeds_many`.
+  Provide embedded functionality to Ecto.
+
+  It provides `one/2` and `many/2` for defining embeds in schemaless changesets.
+
+  ## Struct
+
+  This The embedding struct for `embeds_one` and `embeds_many`.
+
+  This module also provides `one/2` and `many/2` for defining embeds in
+  schemaless changesets.
 
   Its fields are:
 
@@ -33,6 +42,63 @@ defmodule Ecto.Embedded do
     ordered: true
   ]
 
+  @doc """
+  Defines an embeds-one type for a schemaless changeset.
+
+  The field value may be `nil` when there is no existing embed or a struct of
+  the related schema when there is one. Schemaless changesets do not initialize
+  default values, so the data should explicitly contain `nil` for an empty
+  embed. The returned type can be passed in the types map accepted by
+  `Ecto.Changeset.cast/4` and then cast with `Ecto.Changeset.cast_embed/3`:
+
+      types = %{profile: Ecto.Embedded.one(Profile)}
+
+      {%{profile: nil}, types}
+      |> Ecto.Changeset.cast(params, [])
+      |> Ecto.Changeset.cast_embed(:profile)
+
+  The supported option is `:on_replace`, with the same values supported by
+  `Ecto.Schema.embeds_one/3`.
+  """
+  @spec one(module(), keyword()) :: {:embed, %Embedded{}}
+  def one(related, opts \\ []) when is_atom(related) do
+    init_schemaless(:one, related, opts)
+  end
+
+  @doc """
+  Defines an embeds-many type for a schemaless changeset.
+
+  The field value may be an empty list or a list of structs of the related
+  schema. Schemaless changesets do not initialize default values, so the data
+  should explicitly contain an empty list when there are no existing embeds.
+  The returned type can be passed in the types map accepted by
+  `Ecto.Changeset.cast/4` and then cast with `Ecto.Changeset.cast_embed/3`:
+
+      types = %{posts: Ecto.Embedded.many(Post)}
+
+      {%{posts: []}, types}
+      |> Ecto.Changeset.cast(params, [])
+      |> Ecto.Changeset.cast_embed(:posts)
+
+  The supported option is `:on_replace`, with the same values supported by
+  `Ecto.Schema.embeds_many/3`.
+  """
+  @spec many(module(), keyword()) :: {:embed, %Embedded{}}
+  def many(related, opts \\ []) when is_atom(related) do
+    init_schemaless(:many, related, opts)
+  end
+
+  defp init_schemaless(cardinality, related, opts) do
+    case Keyword.keys(opts) -- [:on_replace] do
+      [] ->
+        opts = [cardinality: cardinality, related: related] ++ opts
+        {:embed, init(opts)}
+
+      [option | _] ->
+        raise ArgumentError, "invalid option #{inspect(option)} for #{cardinality}/2"
+    end
+  end
+
   ## Parameterized API
 
   # We treat even embed_many as maps, as that's often the
@@ -49,8 +115,10 @@ defmodule Ecto.Embedded do
       if cardinality == :one, do: @embeds_one_on_replace_opts, else: @on_replace_opts
 
     unless opts[:on_replace] in on_replace_opts do
+      field = if field = opts[:field], do: " for #{inspect(field)}", else: ""
+
       raise ArgumentError,
-            "invalid `:on_replace` option for #{inspect(Keyword.fetch!(opts, :field))}. " <>
+            "invalid `:on_replace` option#{field}. " <>
               "The only valid options are: " <>
               Enum.map_join(on_replace_opts, ", ", &"`#{inspect(&1)}`")
     end
diff --git a/lib/ecto/schema.ex b/lib/ecto/schema.ex
index b293e11e5b..6e9e3ce983 100644
--- a/lib/ecto/schema.ex
+++ b/lib/ecto/schema.ex
@@ -1600,9 +1600,8 @@ defmodule Ecto.Schema do
   @doc ~S"""
   Indicates an embedding of a schema.
 
-  The current schema has zero or one records of the other schema embedded
-  inside of it. It uses a field similar to the `:map` type for storage,
-  but allows embeds to have all the things regular schema can.
+  The current schema keeps zero or a single record of the embedded schema
+  directly inside of it.
 
   You must declare your `embeds_one/3` field with type `:map` at the
   database level.
@@ -1782,8 +1781,8 @@ defmodule Ecto.Schema do
   @doc ~S"""
   Indicates an embedding of many schemas.
 
-  The current schema has zero or more records of the other schema embedded
-  inside of it. Embeds have all the things regular schemas have.
+  The current schema keeps zero or more records of the embedded schema
+  directly inside of it.
 
   It is recommended to declare your `embeds_many/3` field with type `:map`
   in your migrations, instead of using `{:array, :map}`. Ecto can work with
@@ -2117,21 +2116,24 @@ defmodule Ecto.Schema do
     if inserted_at do
       opts = if source = timestamps[:inserted_at_source], do: [source: source], else: []
 
-      opts = if writable = timestamps[:inserted_at_writable] do
-        if writable == :never do
-          raise ArgumentError, ":inserted_at_writable option cannot be set to :never as `inserted_at` will never be populated"
-        end
+      opts =
+        if writable = timestamps[:inserted_at_writable] do
+          if writable == :never do
+            raise ArgumentError,
+                  ":inserted_at_writable option cannot be set to :never as `inserted_at` will never be populated"
+          end
 
-        Keyword.put(opts, :writable, writable)
-      else
-        opts
-      end
+          Keyword.put(opts, :writable, writable)
+        else
+          opts
+        end
 
-      opts = if on_writable_violation = timestamps[:inserted_at_on_writable_violation] do
-        Keyword.put(opts, :on_writable_violation, on_writable_violation)
-      else
-        opts
-      end
+      opts =
+        if on_writable_violation = timestamps[:inserted_at_on_writable_violation] do
+          Keyword.put(opts, :on_writable_violation, on_writable_violation)
+        else
+          opts
+        end
 
       Ecto.Schema.__field__(mod, inserted_at, type, opts)
     end
diff --git a/test/ecto/changeset/embedded_test.exs b/test/ecto/changeset/embedded_test.exs
index fe467aaf1e..ec5e7f5bf3 100644
--- a/test/ecto/changeset/embedded_test.exs
+++ b/test/ecto/changeset/embedded_test.exs
@@ -98,6 +98,10 @@ defmodule Ecto.Changeset.EmbeddedTest do
     end
   end
 
+  defmodule SchemalessAuthor do
+    defstruct profile: nil, posts: []
+  end
+
   defmodule Nested do
     use Ecto.Schema
     import Ecto.Changeset
@@ -179,6 +183,70 @@ defmodule Ecto.Changeset.EmbeddedTest do
     |> Changeset.cast_embed(embed, opts)
   end
 
+  describe "schemaless embeds" do
+    test "casts embeds_one" do
+      types = %{profile: Embedded.one(Profile)}
+
+      changeset =
+        {%SchemalessAuthor{}, types}
+        |> Changeset.cast(%{"profile" => %{"name" => "michal"}}, [])
+        |> Changeset.cast_embed(:profile)
+
+      assert %{profile: profile} = changeset.changes
+      assert profile.data == %Profile{}
+      assert profile.changes == %{name: "michal"}
+      assert profile.action == :insert
+      assert changeset.valid?
+
+      assert {:ok, %SchemalessAuthor{profile: %Profile{name: "michal"}}} =
+               Changeset.apply_action(changeset, :insert)
+
+      assert {:embed, %Embedded{cardinality: :one, field: :profile, owner: SchemalessAuthor}} =
+               changeset.types.profile
+    end
+
+    test "casts embeds_many and preserves nested errors" do
+      types = %{posts: Embedded.many(Post, on_replace: :delete)}
+
+      changeset =
+        {%{posts: []}, types}
+        |> Changeset.cast(
+          %{"posts" => [%{"title" => "valid"}, %{"title" => "no"}]},
+          []
+        )
+        |> Changeset.cast_embed(:posts)
+
+      assert [valid, invalid] = changeset.changes.posts
+      assert valid.changes == %{title: "valid"}
+
+      assert invalid.errors == [
+               title:
+                 {"should be at least %{count} character(s)",
+                  [count: 3, validation: :length, kind: :min, type: :string]}
+             ]
+
+      refute changeset.valid?
+
+      assert {:embed,
+              %Embedded{
+                cardinality: :many,
+                field: :posts,
+                owner: nil,
+                on_replace: :delete
+              }} = changeset.types.posts
+
+      changeset =
+        {%SchemalessAuthor{posts: [%Post{title: "before"}]}, types}
+        |> Changeset.cast(%{"posts" => [%{"title" => "after"}]}, [])
+        |> Changeset.cast_embed(:posts)
+
+      assert [%Changeset{action: :update}] = changeset.changes.posts
+
+      assert {:ok, %SchemalessAuthor{posts: [%Post{title: "after"}]}} =
+               Changeset.apply_action(changeset, :update)
+    end
+  end
+
   ## Cast embeds one
 
   test "cast embeds_one with valid params" do