diff --git a/mautrix_telegram/commands/telegram/misc.py b/mautrix_telegram/commands/telegram/misc.py
index f21f12ce..74ba53c8 100644
--- a/mautrix_telegram/commands/telegram/misc.py
+++ b/mautrix_telegram/commands/telegram/misc.py
@@ -18,6 +18,7 @@ from __future__ import annotations
 from typing import cast
 import base64
 import codecs
+import math
 import re
 
 from aiohttp import ClientSession, InvalidURL
@@ -427,6 +428,9 @@ async def backfill(evt: CommandEvent) -> None:
     if not evt.is_portal:
         await evt.reply("You can only use backfill in portal rooms")
         return
+    elif not evt.config["bridge.backfill.enable"]:
+        await evt.reply("Backfilling is disabled in the bridge config")
+        return
     try:
         limit = int(evt.args[0])
     except (ValueError, IndexError):
@@ -435,16 +439,14 @@ async def backfill(evt: CommandEvent) -> None:
     if not evt.config["bridge.backfill.normal_groups"] and portal.peer_type == "chat":
         await evt.reply("Backfilling normal groups is disabled in the bridge config")
         return
-    try:
-        await portal.backfill(evt.sender, limit=limit)
-    except TakeoutInitDelayError:
-        msg = (
-            "Please accept the data export request from a mobile device, "
-            "then re-run the backfill command."
-        )
-        if portal.peer_type == "user":
-            from mautrix.appservice import IntentAPI
-
-            await portal.main_intent.send_notice(evt.room_id, msg)
-        else:
-            await evt.reply(msg)
+    if portal.backfill_msc2716:
+        messages_per_batch = evt.config["bridge.backfill.incremental.messages_per_batch"]
+        batches = math.ceil(limit / messages_per_batch)
+        rounded = ""
+        if batches * messages_per_batch != limit:
+            rounded = f" (rounded message limit to {batches}*{messages_per_batch})"
+        await portal.enqueue_backfill(evt.sender, priority=0, max_batches=batches)
+        await evt.reply(f"Backfill queued{rounded}")
+    else:
+        output = await portal.forward_backfill(evt.sender, initial=False, override_limit=limit)
+        await evt.reply(output)
diff --git a/mautrix_telegram/config.py b/mautrix_telegram/config.py
index ba225d20..0d66c646 100644
--- a/mautrix_telegram/config.py
+++ b/mautrix_telegram/config.py
@@ -158,12 +158,18 @@ class Config(BaseBridgeConfig):
         copy("bridge.bridge_matrix_leave")
         copy("bridge.kick_on_logout")
         copy("bridge.always_read_joined_telegram_notice")
-        copy("bridge.backfill.invite_own_puppet")
-        copy("bridge.backfill.takeout_limit")
-        copy("bridge.backfill.initial_limit")
-        copy("bridge.backfill.missed_limit")
-        copy("bridge.backfill.disable_notifications")
+        copy("bridge.backfill.enable")
+        copy("bridge.backfill.msc2716")
+        copy("bridge.backfill.double_puppet_backfill")
         copy("bridge.backfill.normal_groups")
+        copy("bridge.backfill.forward.initial_limit")
+        copy("bridge.backfill.forward.sync_limit")
+        copy("bridge.backfill.incremental.messages_per_batch")
+        copy("bridge.backfill.incremental.post_batch_delay")
+        copy("bridge.backfill.incremental.max_batches.user")
+        copy("bridge.backfill.incremental.max_batches.normal_group")
+        copy("bridge.backfill.incremental.max_batches.supergroup")
+        copy("bridge.backfill.incremental.max_batches.channel")
 
         copy("bridge.initial_power_level_overrides.group")
         copy("bridge.initial_power_level_overrides.user")
diff --git a/mautrix_telegram/db/__init__.py b/mautrix_telegram/db/__init__.py
index 40f0706e..482f7867 100644
--- a/mautrix_telegram/db/__init__.py
+++ b/mautrix_telegram/db/__init__.py
@@ -15,6 +15,7 @@
 # along with this program.  If not, see <https://www.gnu.org/licenses/>.
 from mautrix.util.async_db import Database
 
+from .backfill_queue import Backfill
 from .bot_chat import BotChat
 from .disappearing_message import DisappearingMessage
 from .message import Message
@@ -38,6 +39,7 @@ def init(db: Database) -> None:
         BotChat,
         PgSession,
         DisappearingMessage,
+        Backfill,
     ):
         table.db = db
 
@@ -54,4 +56,5 @@ __all__ = [
     "BotChat",
     "PgSession",
     "DisappearingMessage",
+    "Backfill",
 ]
diff --git a/mautrix_telegram/db/backfill_queue.py b/mautrix_telegram/db/backfill_queue.py
new file mode 100644
index 00000000..e86c5def
--- /dev/null
+++ b/mautrix_telegram/db/backfill_queue.py
@@ -0,0 +1,175 @@
+# mautrix-telegram - A Matrix-Telegram puppeting bridge
+# Copyright (C) 2022 Tulir Asokan, Sumner Evans
+#
+# This program is free software: you can redistribute it and/or modify
+# it under the terms of the GNU Affero General Public License as published by
+# the Free Software Foundation, either version 3 of the License, or
+# (at your option) any later version.
+#
+# This program is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+# GNU Affero General Public License for more details.
+#
+# You should have received a copy of the GNU Affero General Public License
+# along with this program.  If not, see <https://www.gnu.org/licenses/>.
+from __future__ import annotations
+
+from typing import TYPE_CHECKING, ClassVar
+from datetime import datetime, timedelta
+
+from asyncpg import Record
+from attr import dataclass
+
+from mautrix.types import RoomID, UserID
+from mautrix.util.async_db import Database
+
+fake_db = Database.create("") if TYPE_CHECKING else None
+
+
+@dataclass
+class Backfill:
+    db: ClassVar[Database] = fake_db
+
+    queue_id: int | None
+    user_mxid: UserID
+    priority: int
+    portal_tgid: int
+    portal_tg_receiver: int
+    messages_per_batch: int
+    post_batch_delay: int
+    max_batches: int
+    dispatch_time: datetime | None
+    completed_at: datetime | None
+    cooldown_timeout: datetime | None
+
+    @staticmethod
+    def new(
+        user_mxid: UserID,
+        priority: int,
+        portal_tgid: int,
+        portal_tg_receiver: int,
+        messages_per_batch: int,
+        post_batch_delay: int = 0,
+        max_batches: int = -1,
+    ) -> "Backfill":
+        return Backfill(
+            queue_id=None,
+            user_mxid=user_mxid,
+            priority=priority,
+            portal_tgid=portal_tgid,
+            portal_tg_receiver=portal_tg_receiver,
+            messages_per_batch=messages_per_batch,
+            post_batch_delay=post_batch_delay,
+            max_batches=max_batches,
+            dispatch_time=None,
+            completed_at=None,
+            cooldown_timeout=None,
+        )
+
+    @classmethod
+    def _from_row(cls, row: Record | None) -> Backfill | None:
+        if row is None:
+            return None
+        return cls(**row)
+
+    columns = [
+        "user_mxid",
+        "priority",
+        "portal_tgid",
+        "portal_tg_receiver",
+        "messages_per_batch",
+        "post_batch_delay",
+        "max_batches",
+        "dispatch_time",
+        "completed_at",
+        "cooldown_timeout",
+    ]
+    columns_str = ",".join(columns)
+
+    @classmethod
+    async def get_next(cls, user_mxid: UserID) -> Backfill | None:
+        q = f"""
+        SELECT queue_id, {cls.columns_str}
+        FROM backfill_queue
+        WHERE user_mxid=$1
+            AND (
+                dispatch_time IS NULL
+                OR (
+                    dispatch_time < $2
+                    AND completed_at IS NULL
+                )
+            )
+            AND (
+                cooldown_timeout IS NULL
+                OR cooldown_timeout < current_timestamp
+            )
+        ORDER BY priority, queue_id
+        LIMIT 1
+        """
+        return cls._from_row(
+            await cls.db.fetchrow(q, user_mxid, datetime.now() - timedelta(minutes=15))
+        )
+
+    @classmethod
+    async def get(
+        cls,
+        user_mxid: UserID,
+        portal_tgid: int,
+        portal_tg_receiver: int,
+    ) -> Backfill | None:
+        q = f"""
+        SELECT queue_id, {cls.columns_str}
+        FROM backfill_queue
+        WHERE user_mxid=$1
+          AND portal_tgid=$2
+          AND portal_tg_receiver=$3
+        ORDER BY priority, queue_id
+        LIMIT 1
+        """
+        return cls._from_row(await cls.db.fetchrow(q, user_mxid, portal_tgid, portal_tg_receiver))
+
+    @classmethod
+    async def delete_all(cls, user_mxid: UserID) -> None:
+        await cls.db.execute("DELETE FROM backfill_queue WHERE user_mxid=$1", user_mxid)
+
+    @classmethod
+    async def delete_for_portal(cls, tgid: int, tg_receiver: int) -> None:
+        q = "DELETE FROM backfill_queue WHERE portal_tgid=$1 AND portal_tg_receiver=$2"
+        await cls.db.execute(q, tgid, tg_receiver)
+
+    async def insert(self) -> None:
+        q = f"""
+        INSERT INTO backfill_queue ({self.columns_str})
+        VALUES ({','.join(f'${i+1}' for i in range(len(self.columns)))})
+        RETURNING queue_id
+        """
+        row = await self.db.fetchrow(
+            q,
+            self.user_mxid,
+            self.priority,
+            self.portal_tgid,
+            self.portal_tg_receiver,
+            self.messages_per_batch,
+            self.post_batch_delay,
+            self.max_batches,
+            self.dispatch_time,
+            self.completed_at,
+            self.cooldown_timeout,
+        )
+        self.queue_id = row["queue_id"]
+
+    async def mark_dispatched(self) -> None:
+        q = "UPDATE backfill_queue SET dispatch_time=$1 WHERE queue_id=$2"
+        await self.db.execute(q, datetime.now(), self.queue_id)
+
+    async def mark_done(self) -> None:
+        q = "UPDATE backfill_queue SET completed_at=$1 WHERE queue_id=$2"
+        await self.db.execute(q, datetime.now(), self.queue_id)
+
+    async def set_cooldown_timeout(self, timeout) -> None:
+        """
+        Set the backfill request to cooldown for ``timeout`` seconds.
+        """
+        q = "UPDATE backfill_queue SET cooldown_timeout=$1 WHERE queue_id=$2"
+        await self.db.execute(q, datetime.now() + timedelta(seconds=timeout), self.queue_id)
diff --git a/mautrix_telegram/db/message.py b/mautrix_telegram/db/message.py
index ae6c3c26..d30376f9 100644
--- a/mautrix_telegram/db/message.py
+++ b/mautrix_telegram/db/message.py
@@ -19,6 +19,7 @@ from typing import TYPE_CHECKING, ClassVar
 
 from asyncpg import Record
 from attr import dataclass
+import attr
 
 from mautrix.types import EventID, RoomID, UserID
 from mautrix.util.async_db import Database, Scheme
@@ -122,6 +123,14 @@ class Message:
         )
         return cls._from_row(await cls.db.fetchrow(q, mx_room, tg_space))
 
+    @classmethod
+    async def find_first(cls, mx_room: RoomID, tg_space: TelegramID) -> Message | None:
+        q = (
+            f"SELECT {cls.columns} FROM message WHERE mx_room=$1 AND tg_space=$2 "
+            f"ORDER BY tgid ASC LIMIT 1"
+        )
+        return cls._from_row(await cls.db.fetchrow(q, mx_room, tg_space))
+
     @classmethod
     async def delete_all(cls, mx_room: RoomID) -> None:
         await cls.db.execute("DELETE FROM message WHERE mx_room=$1", mx_room)
@@ -173,6 +182,23 @@ class Message:
         q = "DELETE FROM message WHERE mxid=$1 AND mx_room=$2"
         await cls.db.execute(q, temp_mxid, mx_room)
 
+    @classmethod
+    async def bulk_insert(cls, messages: list[Message]) -> None:
+        columns = cls.columns.split(", ")
+        records = [attr.astuple(message) for message in messages]
+        async with cls.db.acquire() as conn, conn.transaction():
+            if cls.db.scheme == Scheme.POSTGRES:
+                await conn.copy_records_to_table("message", records=records, columns=columns)
+            else:
+                await conn.executemany(cls._insert_query, records)
+
+    _insert_query: ClassVar[
+        str
+    ] = """
+        INSERT INTO message (mxid, mx_room, tgid, tg_space, edit_index, redacted, content_hash, sender_mxid, sender)
+        VALUES ($1, $2, $3, $4, $5, $6, $7, $8, $9)
+    """
+
     @property
     def _values(self):
         return (
@@ -188,13 +214,7 @@ class Message:
         )
 
     async def insert(self) -> None:
-        q = """
-            INSERT INTO message (
-                mxid, mx_room, tgid, tg_space, edit_index, redacted, content_hash,
-                sender_mxid, sender
-            ) VALUES ($1, $2, $3, $4, $5, $6, $7, $8, $9)
-        """
-        await self.db.execute(q, *self._values)
+        await self.db.execute(self._insert_query, *self._values)
 
     async def delete(self) -> None:
         q = "DELETE FROM message WHERE mxid=$1 AND mx_room=$2 AND tg_space=$3"
diff --git a/mautrix_telegram/example-config.yaml b/mautrix_telegram/example-config.yaml
index 5ec4c552..6937327c 100644
--- a/mautrix_telegram/example-config.yaml
+++ b/mautrix_telegram/example-config.yaml
@@ -337,34 +337,49 @@ bridge:
     create_group_on_invite: true
     # Settings for backfilling messages from Telegram.
     backfill:
-        # Whether or not the Telegram ghosts of logged in Matrix users should be
-        # invited to private chats when backfilling history from Telegram. This is
-        # usually needed to prevent rate limits and to allow timestamp massaging.
-        invite_own_puppet: true
-        # Maximum number of messages to backfill without using a takeout.
-        # The first time a takeout is used, the user has to manually approve it from a different
-        # device. If initial_limit or missed_limit are higher than this value, the bridge will ask
-        # the user to accept the takeout after logging in before syncing any chats.
-        takeout_limit: 100
-        # Maximum number of messages to backfill initially.
-        # Set to 0 to disable backfilling when creating portal, or -1 to disable the limit.
+        # Allow backfilling at all?
+        enable: true
+        # Use MSC2716 for backfilling?
         #
-        # N.B. Initial backfill will only start after member sync. Make sure your
-        #      max_initial_member_sync is set to a low enough value so it doesn't take forever.
-        initial_limit: 0
-        # Maximum number of messages to backfill if messages were missed while the bridge was
-        # disconnected. Note that this only works for logged in users and only if the chat isn't
-        # older than sync_update_limit
-        # Set to 0 to disable backfilling missed messages.
-        missed_limit: 50
-        # If using double puppeting, should notifications be disabled
-        # while the initial backfill is in progress?
-        disable_notifications: false
+        # This requires a server with MSC2716 support, which is currently an experimental feature in Synapse.
+        # It can be enabled by setting experimental_features -> msc2716_enabled to true in homeserver.yaml.
+        msc2716: false
+        # Use double puppets for backfilling?
+        #
+        # If using MSC2716, the double puppets must be in the appservice's user ID namespace
+        # (because the bridge can't use the double puppet access token with batch sending).
+        double_puppet_backfill: false
         # Whether or not to enable backfilling in normal groups.
         # Normal groups have numerous technical problems in Telegram, and backfilling normal groups
         # will likely cause problems if there are multiple Matrix users in the group.
         normal_groups: false
 
+        # Forward backfilling limits. These apply to both MSC2716 and legacy backfill.
+        forward:
+            # Number of messages to backfill immediately after creating a portal.
+            initial_limit: 10
+            # Number of messages to backfill when syncing chats.
+            sync_limit: 100
+
+        # Settings for incremental backfill of history. These only apply when using MSC2716.
+        incremental:
+            # Maximum number of messages to backfill per batch.
+            messages_per_batch: 100
+            # The number of seconds to wait after backfilling the batch of messages.
+            post_batch_delay: 20
+            # The maximum number of batches to backfill per portal, split by the chat type.
+            # If set to -1, all messages in the chat will eventually be backfilled.
+            max_batches:
+                # Direct chats
+                user: -1
+                # Normal groups. Note that the normal_groups option above must be enabled
+                # for these to be backfilled.
+                normal_group: -1
+                # Supergroups
+                supergroup: 10
+                # Broadcast channels
+                channel: -1
+
     # Overrides for base power levels.
     initial_power_level_overrides:
         user: {}
diff --git a/mautrix_telegram/matrix.py b/mautrix_telegram/matrix.py
index 08a741f4..3c163d58 100644
--- a/mautrix_telegram/matrix.py
+++ b/mautrix_telegram/matrix.py
@@ -16,15 +16,14 @@
 from __future__ import annotations
 
 from typing import TYPE_CHECKING
+import sys
 
 from mautrix.bridge import BaseMatrixHandler
-from mautrix.errors import MatrixError
 from mautrix.types import (
     Event,
     EventID,
     EventType,
     MemberStateEventContent,
-    MessageType,
     PresenceEvent,
     PresenceState,
     ReactionEvent,
@@ -36,7 +35,6 @@ from mautrix.types import (
     RoomTopicStateEventContent as TopicContent,
     SingleReceiptEventContent,
     StateEvent,
-    TextMessageEventContent,
     TypingEvent,
     UserID,
 )
@@ -63,6 +61,21 @@ class MatrixHandler(BaseMatrixHandler):
 
         self._previously_typing = {}
 
+    async def check_versions(self) -> None:
+        await super().check_versions()
+        if self.config["bridge.backfill.msc2716"] and not (
+            support := self.versions.supports("org.matrix.msc2716")
+        ):
+            self.log.fatal(
+                "Backfilling with MSC2716 is enabled in bridge config, but "
+                + (
+                    "batch sending is not enabled on homeserver"
+                    if support is False
+                    else "homeserver does not support batch sending"
+                )
+            )
+            sys.exit(18)
+
     async def handle_puppet_group_invite(
         self,
         room_id: RoomID,
diff --git a/mautrix_telegram/portal.py b/mautrix_telegram/portal.py
index 9e6ab0cf..3d883c1d 100644
--- a/mautrix_telegram/portal.py
+++ b/mautrix_telegram/portal.py
@@ -132,12 +132,15 @@ from telethon.tl.types import (
     UserProfilePhotoEmpty,
 )
 from telethon.utils import encode_waveform
+import attr
 
 from mautrix.appservice import DOUBLE_PUPPET_SOURCE_KEY, IntentAPI
 from mautrix.bridge import BasePortal, NotificationDisabler, RejectMatrixInvite, async_getter_lock
 from mautrix.errors import IntentError, MatrixRequestError, MForbidden
 from mautrix.types import (
     BatchID,
+    BatchSendEvent,
+    BatchSendStateEvent,
     BeeperMessageStatusEventContent,
     ContentURI,
     EventID,
@@ -148,6 +151,7 @@ from mautrix.types import (
     LocationMessageEventContent,
     MediaMessageEventContent,
     Membership,
+    MemberStateEventContent,
     MessageEventContent,
     MessageStatus,
     MessageStatusReason,
@@ -182,6 +186,7 @@ from . import (
 )
 from .config import Config
 from .db import (
+    Backfill,
     DisappearingMessage,
     Message as DBMessage,
     Portal as DBPortal,
@@ -204,6 +209,7 @@ if TYPE_CHECKING:
 StateBridge = EventType.find("m.bridge", EventType.Class.STATE)
 StateHalfShotBridge = EventType.find("uk.half-shot.bridge", EventType.Class.STATE)
 DummyPortalCreated = EventType.find("fi.mau.dummy.portal_created", EventType.Class.MESSAGE)
+StateMarker = EventType.find("org.matrix.msc2716.marker", EventType.Class.STATE)
 
 InviteList = Union[UserID, List[UserID]]
 UpdateTyping = Union[UpdateUserTyping, UpdateChatUserTyping, UpdateChannelUserTyping]
@@ -250,6 +256,7 @@ class Portal(DBPortal, BasePortal):
     backfill_lock: SimpleLock
     backfill_method_lock: asyncio.Lock
     backfill_leave: set[IntentAPI] | None
+    backfill_msc2716: bool
 
     alias: RoomAlias | None
 
@@ -269,6 +276,7 @@ class Portal(DBPortal, BasePortal):
     _sponsored_seen: dict[UserID, bool]
     _new_messages_after_sponsored: bool
 
+    _member_list_cache: dict[EventID, set[UserID]]
     _prev_reaction_poll: dict[UserID, float]
 
     _msg_conv: putil.TelegramMessageConverter
@@ -322,6 +330,7 @@ class Portal(DBPortal, BasePortal):
         self.log = self.log.getChild(self.tgid_log if self.tgid else self.mxid)
         self._main_intent = None
         self.deleted = False
+
         self.backfill_lock = SimpleLock(
             "Waiting for backfilling to finish before handling %s", log=self.log
         )
@@ -341,6 +350,7 @@ class Portal(DBPortal, BasePortal):
         self._new_messages_after_sponsored = True
         self._bridging_blocked_at_runtime = False
 
+        self._member_list_cache = {}
         self._prev_reaction_poll = defaultdict(lambda: 0.0)
 
         self._msg_conv = putil.TelegramMessageConverter(self)
@@ -428,6 +438,7 @@ class Portal(DBPortal, BasePortal):
         cls.filter_mode = cls.config["bridge.filter.mode"]
         cls.filter_list = cls.config["bridge.filter.list"]
         cls.hs_domain = cls.config["homeserver.domain"]
+        cls.backfill_msc2716 = cls.config["bridge.backfill.msc2716"]
         cls.alias_template = SimpleTemplate(
             cls.config["bridge.alias_template"],
             "groupname",
@@ -905,32 +916,30 @@ class Portal(DBPortal, BasePortal):
 
             self.mxid = room_id
             self.by_mxid[self.mxid] = self
-            self.first_event_id = await self.main_intent.send_message_event(
-                self.mxid, DummyPortalCreated, {}
-            )
             await self.save()
             self.log.debug(f"Matrix room created: {self.mxid}")
             await self.az.state_store.set_power_levels(self.mxid, power_levels)
             await user.register_portal(self)
 
             await self.invite_to_matrix(invites)
+            await self.update_matrix_room(user, entity, puppet, levels=power_levels, users=users)
 
-            update_room = asyncio.create_task(
-                self.update_matrix_room(user, entity, puppet, levels=power_levels, users=users)
+            self.first_event_id = await self.main_intent.send_message_event(
+                self.mxid, DummyPortalCreated, {}
             )
-
-            if self.config["bridge.backfill.initial_limit"] > 0:
-                self.log.debug(
-                    "Initial backfill is enabled. Waiting for room members to sync "
-                    "and then starting backfill"
+            if not self.bridge.homeserver_software.is_hungry:
+                self._member_list_cache[self.first_event_id] = set(
+                    (await self.main_intent.get_joined_members(self.mxid)).keys()
                 )
-                await update_room
+            await self.save()
 
+            if isinstance(user, u.User) or not self.backfill_msc2716:
                 try:
-                    if isinstance(user, u.User):
-                        await self.backfill(user, is_initial=True)
+                    await self.forward_backfill(user, initial=True)
                 except Exception:
-                    self.log.exception("Failed to backfill new portal")
+                    self.log.exception("Error in initial backfill")
+                if self.backfill_msc2716:
+                    await self.enqueue_backfill(user, priority=50)
 
         return self.mxid
 
@@ -2595,142 +2604,188 @@ class Portal(DBPortal, BasePortal):
         await DBMessage.replace_temp_mxid(temporary_identifier, self.mxid, event_id)
 
     @property
-    def _takeout_options(self) -> dict[str, bool | int]:
-        return {
-            "files": True,
-            "megagroups": self.megagroup,
-            "chats": self.peer_type == "chat",
-            "users": self.peer_type == "user",
-            "channels": (self.peer_type == "channel" and not self.megagroup),
-            "max_file_size": min(self.matrix.media_config.upload_size, 2000 * 1024 * 1024),
-        }
+    def _default_max_batches(self) -> int:
+        if self.peer_type == "user":
+            own_type = "user"
+        elif self.peer_type == "chat":
+            own_type = "normal_group"
+        elif self.megagroup:
+            own_type = "supergroup"
+        else:
+            own_type = "channel"
+        return self.config[f"bridge.backfill.incremental.max_batches.{own_type}"]
+
+    async def enqueue_backfill(
+        self,
+        source: u.User,
+        priority: int,
+        max_batches: int | None = None,
+        messages_per_batch: int | None = None,
+    ) -> None:
+        # TODO check that there are no queued backfills
+        # if not await Backfill.get(source.mxid, self.tgid, self.tg_receiver):
+        await Backfill.new(
+            user_mxid=source.mxid,
+            priority=priority,
+            portal_tgid=self.tgid,
+            portal_tg_receiver=self.tg_receiver,
+            messages_per_batch=(
+                messages_per_batch or self.config["bridge.backfill.incremental.messages_per_batch"]
+            ),
+            post_batch_delay=self.config["bridge.backfill.incremental.post_batch_delay"],
+            max_batches=max_batches or self._default_max_batches,
+        ).insert()
+        source.wakeup_backfill_task.set()
+
+    async def forward_backfill(
+        self,
+        source: u.User,
+        initial: bool,
+        last_tgid: int | None = None,
+        override_limit: int | None = None,
+    ) -> str:
+        type = "initial" if initial else "sync"
+        limit = override_limit or self.config[f"bridge.backfill.forward.{type}_limit"]
+        with self.backfill_lock:
+            output = await self.backfill(
+                source, source.client, forward=True, forward_limit=limit, last_tgid=last_tgid
+            )
+            self.log.debug(f"Forward backfill complete, status: {output}")
+            return output
 
     async def backfill(
         self,
         source: u.User,
-        is_initial: bool = False,
-        limit: int | None = None,
-        last_id: int | None = None,
-    ) -> None:
+        client: MautrixTelegramClient,
+        req: Backfill | None = None,
+        forward: bool = False,
+        forward_limit: int | None = None,
+        last_tgid: int | None = None,
+    ) -> str:
         async with self.backfill_method_lock:
-            await self._locked_backfill(source, is_initial, limit, last_id)
+            return await self._locked_backfill(
+                source, client, req, forward, forward_limit, last_tgid
+            )
 
     async def _locked_backfill(
         self,
         source: u.User,
-        is_initial: bool = False,
-        limit: int | None = None,
+        client: MautrixTelegramClient,
+        req: Backfill | None = None,
+        forward: bool = False,
+        forward_limit: int | None = None,
         last_tgid: int | None = None,
-    ) -> None:
-        limit = limit or (
-            self.config["bridge.backfill.initial_limit"]
-            if is_initial
-            else self.config["bridge.backfill.missed_limit"]
-        )
-        if limit == 0:
-            return
+    ) -> str:
+        assert forward != bool(req)
         if not self.config["bridge.backfill.normal_groups"] and self.peer_type == "chat":
-            return
-        last_in_room = await DBMessage.find_last(
-            self.mxid, (source.tgid if self.peer_type != "channel" else self.tgid)
-        )
-        min_id = last_in_room.tgid if last_in_room else 0
-        if last_tgid is None:
-            messages = await source.client.get_messages(self.peer, limit=1)
-            if not messages:
-                # The chat seems empty
-                return
-            last_tgid = messages[0].id
-        if last_tgid <= min_id or (last_tgid == 1 and self.peer_type == "channel"):
-            # Nothing to backfill
-            return
-        if limit < 0:
-            limit = last_tgid - min_id
-            limit_type = "unlimited"
-        elif self.peer_type == "channel":
-            min_id = max(last_tgid - limit, min_id)
-            # This is now just an approximate message count, not the actual limit.
-            limit = last_tgid - min_id
-            limit_type = "channel"
-        else:
-            # This limit will be higher than the actual message count if there are any messages
-            # in other DMs or normal groups, but that's not too bad.
-            limit = min(last_tgid - min_id, limit)
-            limit_type = "dm/minigroup"
-        self.log.debug(
-            f"Backfilling up to {limit} messages after ID {min_id} through {source.mxid} "
-            f"(last message: {last_tgid}, limit type: {limit_type})"
-        )
-        with self.backfill_lock:
-            await self._backfill(source, min_id, limit)
-
-    async def _backfill(self, source: u.User, min_id: int, limit: int) -> None:
-        self.backfill_leave = set()
-        if (
-            self.peer_type == "user"
-            and self.tgid != source.tgid
-            and self.config["bridge.backfill.invite_own_puppet"]
-        ):
-            self.log.debug("Adding %s's default puppet to room for backfilling", source.mxid)
-            sender = await p.Puppet.get_by_tgid(source.tgid)
-            await self.main_intent.invite_user(self.mxid, sender.default_mxid)
-            await sender.default_mxid_intent.join_room_by_id(self.mxid)
-            self.backfill_leave.add(sender.default_mxid_intent)
-
-        client = source.client
-        async with NotificationDisabler(self.mxid, source):
-            if limit > self.config["bridge.backfill.takeout_limit"]:
-                self.log.debug(f"Opening takeout client for {source.tgid}")
-                async with client.takeout(**self._takeout_options) as takeout:
-                    count, handled = await self._backfill_messages(source, min_id, limit, takeout)
+            return "Backfilling normal groups is disabled in the bridge config"
+        tg_space = source.tgid if self.peer_type != "channel" else self.tgid
+        prev_event_id = self.first_event_id
+        if forward:
+            last_in_room = await DBMessage.find_last(self.mxid, tg_space)
+            if last_in_room:
+                prev_event_id = last_in_room.mxid
+                min_id = last_in_room.tgid
             else:
-                count, handled = await self._backfill_messages(source, min_id, limit, client)
-
-        for intent in self.backfill_leave:
-            self.log.trace("Leaving room with %s post-backfill", intent.mxid)
-            await intent.leave_room(self.mxid)
-        self.backfill_leave = None
-        self.log.info(
-            "Backfilled %d (of %d fetched) messages through %s", handled, count, source.mxid
-        )
-
-    async def _backfill_messages(
-        self, source: u.User, min_id: int, limit: int, client: MautrixTelegramClient
-    ) -> tuple[int, int]:
-        count = handled_count = 0
-        entity = await self.get_input_entity(source)
-        if self.peer_type == "channel":
-            # This is a channel or supergroup, so we'll backfill messages based on the ID.
-            # There are some cases, such as deleted messages, where this may backfill less
-            # messages than the limit.
-            self.log.debug(f"Iterating all messages starting with {min_id} (approx: {limit})")
-            messages = client.iter_messages(entity, reverse=True, min_id=min_id)
-            async for message in messages:
-                count += 1
-                was_handled = await self._handle_telegram_backfill_message(source, message)
-                handled_count += 1 if was_handled else 0
-        else:
-            # Private chats and normal groups don't have their own message ID namespace,
-            # which means we'll have to fetch messages a different way.
+                min_id = 0
+            if last_tgid is None:
+                messages = await source.client.get_messages(self.peer, limit=1)
+                if not messages:
+                    return "Chat is empty, nothing to backfill"
+                last_tgid = messages[0].id
+            if last_tgid <= min_id or (last_tgid == 1 and self.peer_type == "channel"):
+                return (
+                    f"Last bridged message {min_id} is equal to or greater than last message "
+                    f"in Telegram chat {last_tgid}, nothing to backfill"
+                )
+            limit = last_tgid - min_id
+            if (forward_limit or 0) > 0:
+                limit = min(limit, forward_limit)
             self.log.debug(
-                f"Fetching up to {limit} most recent messages, ignoring anything before {min_id}"
+                f"Backfilling up to {limit} messages after ID {min_id} through {source.mxid} "
+                f"(last message: {last_tgid})"
             )
-            messages = await client.get_messages(entity, min_id=min_id, limit=limit)
-            for message in reversed(messages):
-                count += 1
-                if message.id <= min_id:
-                    self.log.trace(
-                        f"Skipping {message.id} in backfill response as it's lower than "
-                        f"the last bridged message ({min_id})"
-                    )
-                    continue
-                was_handled = await self._handle_telegram_backfill_message(source, message)
-                handled_count += 1 if was_handled else 0
-        return count, handled_count
+            anchor_id = min_id
+        else:
+            limit = req.messages_per_batch
+            first_in_room = await DBMessage.find_first(self.mxid, tg_space)
+            anchor_id = first_in_room.tgid if first_in_room else None
+            self.log.debug(
+                f"Backfilling up to {req.messages_per_batch} historical messages "
+                f"before {anchor_id} through {source.mxid}"
+            )
+        insertion_id, message_count, first_id = await self._backfill_messages(
+            source, client, forward, anchor_id, limit, prev_event_id
+        )
+        if prev_event_id == self.first_event_id:
+            if insertion_id and not self.base_insertion_id:
+                self.base_insertion_id = insertion_id
+            elif not insertion_id:
+                insertion_id = self.base_insertion_id
+        await self.save()
+        # TODO this should probably check actual event count instead of message count
+        if message_count > 0 and self.backfill_msc2716:
+            await self.main_intent.send_state_event(
+                self.mxid,
+                StateMarker,
+                {
+                    "org.matrix.msc2716.marker.insertion": insertion_id,
+                    "com.beeper.timestamp": int(time.time() * 1000),
+                },
+                state_key=insertion_id,
+            )
+        if forward:
+            self.log.debug(f"Forward backfill finished with {message_count} messages")
+        elif message_count > 0 and first_id and first_id > 1:
+            if req.max_batches in (0, 1):
+                return "Already backfilled enough batches, not enqueuing more"
+            self.log.debug(f"Enqueuing more backfill through {source.mxid}")
+            await self.enqueue_backfill(
+                source,
+                priority=100,
+                messages_per_batch=req.messages_per_batch,
+                max_batches=-1 if req.max_batches < 0 else (req.max_batches - 1),
+            )
+        else:
+            self.log.debug("No more messages to backfill")
+        return f"Backfilled {message_count} messages"
 
-    async def _handle_telegram_backfill_message(
-        self, source: au.AbstractUser, msg: Message | MessageService
-    ) -> bool:
+    def _can_double_puppet_backfill(self, custom_mxid: UserID) -> bool:
+        if not self.backfill_msc2716:
+            return True
+        if not self.config["bridge.backfill.double_puppet_backfill"]:
+            return False
+        if self.bridge.homeserver_software.is_hungry:
+            return True
+
+        # Batch sending can only use local users, so don't allow double puppets on other servers.
+        if custom_mxid[custom_mxid.index(":") + 1 :] != self.config["homeserver.domain"]:
+            return False
+        return True
+
+    async def _get_members_at(self, event_id: EventID) -> set[UserID]:
+        try:
+            return self._member_list_cache[event_id]
+        except KeyError:
+            pass
+        # TODO cache the list in db?
+        self.log.debug(f"Fetching member list at {event_id}")
+        ctx = await self.main_intent.get_event_context(self.mxid, event_id, limit=0)
+        members = {
+            evt.state_key
+            for evt in ctx.state
+            if evt.type == EventType.ROOM_MEMBER and evt.content.membership == Membership.JOIN
+        }
+        self.log.debug(f"Found {len(members)} members at {event_id}")
+        self._member_list_cache[event_id] = members
+        return members
+
+    async def _convert_batch_msg(
+        self,
+        source: u.User,
+        msg: Message,
+        add_member: Callable[[IntentAPI, str, ContentURI], Awaitable[None]],
+    ) -> tuple[putil.ConvertedMessage, IntentAPI]:
         if msg.from_id and isinstance(msg.from_id, (PeerUser, PeerChannel)):
             sender = await p.Puppet.get_by_peer(msg.from_id)
         elif isinstance(msg.peer_id, PeerUser):
@@ -2740,20 +2795,171 @@ class Portal(DBPortal, BasePortal):
                 sender = await p.Puppet.get_by_peer(msg.peer_id)
         else:
             sender = None
-        if isinstance(msg, MessageService):
-            if isinstance(msg.action, MessageActionContactSignUp):
-                await self.handle_telegram_joined(source, sender, msg, backfill=True)
-                return True
-            else:
-                self.log.debug(
-                    f"Unhandled service message {type(msg.action).__name__} in backfill"
-                )
-        elif isinstance(msg, Message):
-            await self.handle_telegram_message(source, sender, msg)
-            return True
+        if sender:
+            intent = sender.intent_for(self)
+            if not sender.displayname:
+                entity = await source.client.get_entity(sender.peer)
+                await sender.update_info(source, entity)
         else:
-            self.log.debug(f"Unhandled message type {type(msg).__name__} in backfill")
-        return False
+            intent = self.main_intent
+        if intent.api.is_real_user and not self._can_double_puppet_backfill(intent.mxid):
+            intent = sender.default_mxid_intent
+        if sender:
+            await add_member(intent, sender.displayname, sender.avatar_url)
+        is_bot = sender.is_bot if sender else False
+        converted = await self._msg_conv.convert(source, intent, is_bot, msg)
+        return converted, intent
+
+    async def _wrap_batch_msg(
+        self,
+        intent: IntentAPI,
+        msg: Message,
+        converted: putil.ConvertedMessage,
+        caption: bool = False,
+    ) -> BatchSendEvent:
+        if caption:
+            content = converted.caption
+            event_type = EventType.ROOM_MESSAGE
+        else:
+            content = converted.content
+            event_type = converted.type
+        if self.encrypted and self.matrix.e2ee:
+            event_type, content = await self.matrix.e2ee.encrypt(self.mxid, event_type, content)
+        return BatchSendEvent(
+            sender=intent.mxid,
+            timestamp=int(msg.date.timestamp() * 1000),
+            content=content,
+            type=event_type,
+        )
+
+    async def _backfill_messages(
+        self,
+        source: u.User,
+        client: MautrixTelegramClient,
+        forward: bool,
+        anchor_id: int,
+        limit: int,
+        prev_event_id: EventID,
+    ) -> tuple[EventID | None, int, int]:
+        entity = await self.get_input_entity(source)
+        events = []
+        intents = []
+        metas = []
+        state_events = []
+        do_batch_send = self.backfill_msc2716
+        added_members = (
+            await self._get_members_at(prev_event_id)
+            if not self.bridge.homeserver_software.is_hungry and do_batch_send
+            else []
+        )
+        before_first_msg_timestamp = 0
+
+        async def add_member(intent: IntentAPI, displayname: str, avatar_url: ContentURI) -> None:
+            if self.bridge.homeserver_software.is_hungry or intent.mxid in added_members:
+                return
+            added_members.add(intent.mxid)
+            if not do_batch_send:
+                # TODO leave these members?
+                await intent.ensure_joined(self.mxid)
+                return
+            invite_event = BatchSendStateEvent(
+                type=EventType.ROOM_MEMBER,
+                state_key=intent.mxid,
+                sender=self.main_intent.mxid,
+                timestamp=before_first_msg_timestamp,
+                content=MemberStateEventContent(
+                    membership=Membership.INVITE,
+                    displayname=displayname,
+                    avatar_url=avatar_url,
+                ),
+            )
+            join_event = attr.evolve(
+                invite_event,
+                content=attr.evolve(invite_event.content, membership=Membership.JOIN),
+                sender=intent.mxid,
+            )
+            state_events.append(invite_event)
+            state_events.append(join_event)
+
+        first_id = anchor_id
+        message_count = 0
+        minmax = {"min_id": anchor_id} if forward else {"max_id": anchor_id}
+        if not forward and not anchor_id:
+            anchor_id = 2**31 - 1
+            minmax = {}
+        self.log.debug(f"Iterating messages through {source.tgid} with {limit=}, {minmax}")
+        # Iterate messages newest to oldest and collect the results
+        async for msg in client.iter_messages(entity, limit=limit, **minmax):
+            message_count += 1
+            if (forward and msg.id <= anchor_id) or (not forward and msg.id >= anchor_id):
+                continue
+            elif isinstance(msg, MessageService):
+                # TODO some service messages can be backfilled
+                continue
+            if not before_first_msg_timestamp:
+                first_id = msg.id
+                before_first_msg_timestamp = int(msg.date.timestamp() * 1000) - 1
+
+            converted, intent = await self._convert_batch_msg(source, msg, add_member)
+            events.append(await self._wrap_batch_msg(intent, msg, converted))
+            intents.append(intent)
+            metas.append(msg)
+            if converted.caption:
+                events.append(await self._wrap_batch_msg(intent, msg, converted, caption=True))
+                intents.append(intent)
+                metas.append(None)
+        if len(events) == 0:
+            self.log.debug(
+                f"Didn't get any events to send out of {message_count} messages fetched "
+                f"(first received ID: {first_id})"
+            )
+            return None, message_count, first_id
+        self.log.debug(
+            f"Got {len(events)} events to send out of {message_count} messages fetched "
+            f"(first received ID: {first_id})"
+        )
+        if do_batch_send:
+            resp = await self.main_intent.batch_send(
+                self.mxid,
+                prev_event_id,
+                batch_id=self.next_batch_id if not forward else None,
+                # We iterated the events in reverse chronological order,
+                # so reverse them before sending
+                events=list(reversed(events)),
+                state_events_at_start=state_events,
+                beeper_new_messages=forward,
+            )
+            if prev_event_id == self.first_event_id and resp.next_batch_id:
+                self.next_batch_id = resp.next_batch_id
+            base_insertion_event_id = resp.base_insertion_event_id
+            event_ids = resp.event_ids
+        else:
+            base_insertion_event_id = None
+            event_ids = [
+                await intent.send_message_event(
+                    self.mxid, evt.type, evt.content, timestamp=evt.timestamp
+                )
+                for evt, intent in zip(reversed(events), reversed(intents))
+            ]
+        tg_space = source.tgid if self.peer_type != "channel" else self.tgid
+        await DBMessage.bulk_insert(
+            [
+                DBMessage(
+                    mxid=event_id,
+                    mx_room=self.mxid,
+                    tgid=msg.id,
+                    tg_space=tg_space,
+                    edit_index=0,
+                    content_hash=self.dedup.hash_event(msg),
+                    # TODO sender
+                )
+                # Original arrays are in reverse chronological order, but event IDs are
+                # chronological (because we reversed the original messages list before sending)
+                for event_id, msg in zip(event_ids, reversed(metas))
+                if msg is not None
+            ]
+        )
+        return base_insertion_event_id, message_count, first_id
 
     def _split_dm_reaction_counts(self, counts: list[ReactionCount]) -> list[MessagePeerReaction]:
         reactions = []
diff --git a/mautrix_telegram/puppet.py b/mautrix_telegram/puppet.py
index 5a34d1cf..47883fbb 100644
--- a/mautrix_telegram/puppet.py
+++ b/mautrix_telegram/puppet.py
@@ -399,7 +399,7 @@ class Puppet(DBPuppet, BasePuppet):
 
     async def default_puppet_should_leave_room(self, room_id: RoomID) -> bool:
         portal: p.Portal = await p.Portal.get_by_mxid(room_id)
-        return portal and not portal.backfill_lock.locked and portal.peer_type != "user"
+        return portal and portal.peer_type != "user"
 
     # endregion
     # region Getters
diff --git a/mautrix_telegram/user.py b/mautrix_telegram/user.py
index c03d8cfd..34537383 100644
--- a/mautrix_telegram/user.py
+++ b/mautrix_telegram/user.py
@@ -20,7 +20,12 @@ from datetime import datetime, timezone
 import asyncio
 import time
 
-from telethon.errors import AuthKeyDuplicatedError, RPCError, UnauthorizedError
+from telethon.errors import (
+    AuthKeyDuplicatedError,
+    RPCError,
+    TakeoutInitDelayError,
+    UnauthorizedError,
+)
 from telethon.tl.custom import Dialog
 from telethon.tl.functions.account import UpdateStatusRequest
 from telethon.tl.functions.contacts import GetContactsRequest, SearchRequest
@@ -57,7 +62,8 @@ from mautrix.util.opt_prometheus import Gauge
 
 from . import portal as po, puppet as pu, util
 from .abstract_user import AbstractUser
-from .db import Message as DBMessage, PgSession, User as DBUser
+from .db import Backfill, Message as DBMessage, PgSession, User as DBUser
+from .tgclient import MautrixTelegramClient
 from .types import TelegramID
 
 if TYPE_CHECKING:
@@ -86,7 +92,11 @@ class User(DBUser, AbstractUser, BaseUser):
 
     _ensure_started_lock: asyncio.Lock
     _track_connection_task: asyncio.Task | None
+    _backfill_task: asyncio.Task | None
+    wakeup_backfill_task: asyncio.Event
     _is_backfilling: bool
+    takeout_retry_immediate: asyncio.Event
+    takeout_requested: bool
 
     _available_emoji_reactions: set[str] | None
     _available_emoji_reactions_hash: int | None
@@ -119,6 +129,12 @@ class User(DBUser, AbstractUser, BaseUser):
         self._track_connection_task = None
         self._is_backfilling = False
         self._portals_cache = None
+
+        self._backfill_task = None
+        self.wakeup_backfill_task = asyncio.Event()
+        self.takeout_retry_immediate = asyncio.Event()
+        self.takeout_requested = False
+
         self._available_emoji_reactions = None
         self._available_emoji_reactions_hash = None
         self._available_emoji_reactions_fetched = 0
@@ -248,6 +264,14 @@ class User(DBUser, AbstractUser, BaseUser):
             self.client and self.client._sender and self.client._sender._transport_connected()
         )
 
+    @property
+    def _bridge_state_info(self) -> dict[str, Any]:
+        if self.takeout_requested:
+            return {
+                "takeout_requested": True,
+            }
+        return {}
+
     async def _track_connection(self) -> None:
         self.log.debug("Starting loop to track connection state")
         while True:
@@ -260,6 +284,7 @@ class User(DBUser, AbstractUser, BaseUser):
                     if self._is_backfilling
                     else BridgeStateEvent.CONNECTED,
                     ttl=3600,
+                    info=self._bridge_state_info,
                 )
             else:
                 await self.push_bridge_state(
@@ -284,7 +309,7 @@ class User(DBUser, AbstractUser, BaseUser):
         else:
             state_event = BridgeStateEvent.UNKNOWN_ERROR
             ttl = 240
-        return [BridgeState(state_event=state_event, ttl=ttl)]
+        return [BridgeState(state_event=state_event, ttl=ttl, info=self._bridge_state_info)]
 
     async def get_puppet(self) -> pu.Puppet | None:
         if not self.tgid:
@@ -302,6 +327,9 @@ class User(DBUser, AbstractUser, BaseUser):
         if self._track_connection_task:
             self._track_connection_task.cancel()
             self._track_connection_task = None
+        if self._backfill_task:
+            self._backfill_task.cancel()
+            self._backfill_task = None
         await super().stop()
         self._track_metric(METRIC_CONNECTED, False)
 
@@ -318,6 +346,8 @@ class User(DBUser, AbstractUser, BaseUser):
             return
 
         self._track_metric(METRIC_LOGGED_IN, True)
+        if not self._backfill_task or self._backfill_task.done():
+            self._backfill_task = asyncio.create_task(self._handle_backfill_requests_loop())
 
         try:
             puppet = await pu.Puppet.get_by_tgid(self.tgid)
@@ -327,7 +357,7 @@ class User(DBUser, AbstractUser, BaseUser):
         except Exception:
             self.log.exception("Failed to automatically enable custom puppet")
 
-        if not self.is_bot and self.config["bridge.startup_sync"]:
+        if not self.is_bot and (self.config["bridge.startup_sync"] or first_login):
             try:
                 self._is_backfilling = True
                 await self.sync_dialogs()
@@ -337,6 +367,80 @@ class User(DBUser, AbstractUser, BaseUser):
             finally:
                 self._is_backfilling = False
 
+    @property
+    def _takeout_options(self) -> dict[str, bool | int]:
+        return {
+            "users": True,
+            "chats": self.config["bridge.backfill.normal_groups"],
+            "megagroups": True,
+            "channels": True,
+            "files": True,
+            "max_file_size": min(self.bridge.matrix.media_config.upload_size, 2000 * 1024 * 1024),
+        }
+
+    async def _handle_backfill_requests_loop(self) -> None:
+        while True:
+            req = await Backfill.get_next(self.mxid)
+            if not req:
+                await self.wakeup_backfill_task.wait()
+                self.wakeup_backfill_task.clear()
+            else:
+                await self._takeout_and_backfill(req)
+
+    async def _takeout_and_backfill(self, first_req: Backfill, first_attempt: bool = True) -> None:
+        self.takeout_retry_immediate.clear()
+        self.takeout_requested = True
+        try:
+            async with self.client.takeout(**self._takeout_options) as takeout_client:
+                self.takeout_requested = False
+                self.log.info("Acquired takeout client successfully")
+                await self._backfill_loop_with_client(takeout_client, first_req)
+                self.log.info("Backfills finished, exiting takeout")
+        except TakeoutInitDelayError as e:
+            if first_attempt:
+                self.log.info(
+                    f"Takeout requested, will wait for retry request or {e.seconds} seconds"
+                )
+            else:
+                self.log.warning(
+                    f"Got takeout init delay again after retry, waiting for {e.seconds} seconds"
+                )
+            try:
+                await asyncio.wait_for(self.takeout_retry_immediate.wait(), timeout=e.seconds)
+                self.log.info("Retrying takeout")
+            except asyncio.TimeoutError:
+                self.log.info("Takeout timeout expired")
+            await self._takeout_and_backfill(first_req, first_attempt=False)
+
+    async def _backfill_loop_with_client(
+        self, client: MautrixTelegramClient, first_req: Backfill
+    ) -> None:
+        missed_reqs = 0
+        while missed_reqs < 20:
+            req = first_req or await Backfill.get_next(self.mxid)
+            first_req = None
+            if not req:
+                missed_reqs += 1
+                try:
+                    await asyncio.wait_for(self.wakeup_backfill_task.wait(), timeout=30)
+                except asyncio.TimeoutError:
+                    pass
+                self.wakeup_backfill_task.clear()
+                continue
+            missed_reqs = 0
+            self.log.info("Backfill request %s", req)
+            try:
+                portal = await po.Portal.get_by_tgid(
+                    TelegramID(req.portal_tgid), tg_receiver=TelegramID(req.portal_tg_receiver)
+                )
+                await req.mark_dispatched()
+                await portal.backfill(self, client, req=req)
+                await req.mark_done()
+                await asyncio.sleep(req.post_batch_delay)
+            except Exception:
+                self.log.exception("Error handling backfill request for %s", req.portal_tgid)
+                await req.set_cooldown_timeout(10)
+
     async def update(self, update: TypeUpdate) -> bool:
         if not self.is_bot:
             return False
@@ -573,7 +677,7 @@ class User(DBUser, AbstractUser, BaseUser):
         was_created = False
         if portal.mxid:
             try:
-                await portal.backfill(self, last_id=dialog.message.id)
+                await portal.forward_backfill(self, initial=False, last_tgid=dialog.message.id)
             except Exception:
                 self.log.exception(f"Error while backfilling {portal.tgid_log}")
             try:
diff --git a/mautrix_telegram/web/provisioning/__init__.py b/mautrix_telegram/web/provisioning/__init__.py
index 3217529b..006c8911 100644
--- a/mautrix_telegram/web/provisioning/__init__.py
+++ b/mautrix_telegram/web/provisioning/__init__.py
@@ -71,6 +71,8 @@ class ProvisioningAPI(AuthAPI):
         )
         self.app.router.add_route("POST", f"{user_prefix}/pm/{{identifier}}", self.start_dm)
 
+        self.app.router.add_route("POST", f"{user_prefix}/retry_takeout", self.retry_takeout)
+
         self.app.router.add_route("POST", f"{user_prefix}/logout", self.logout)
         self.app.router.add_route("POST", f"{user_prefix}/login/bot_token", self.send_bot_token)
         self.app.router.add_route("POST", f"{user_prefix}/login/request_code", self.request_code)
@@ -494,6 +496,22 @@ class ProvisioningAPI(AuthAPI):
             status=201 if just_created else 200,
         )
 
+    async def retry_takeout(self, request: web.Request) -> web.Response:
+        data, user, err = await self.get_user_request_info(
+            request, expect_logged_in=True, want_data=False
+        )
+        if err is not None:
+            return err
+        if not user.takeout_requested:
+            return web.json_response(
+                {
+                    "error": "There was no takeout requested",
+                },
+                status=400,
+            )
+        user.takeout_retry_immediate.set()
+        return web.json_response({}, status=200)
+
     async def send_bot_token(self, request: web.Request) -> web.Response:
         data, user, err = await self.get_user_request_info(request)
         if err is not None:
diff --git a/requirements.txt b/requirements.txt
index 86f3281e..1971bf3b 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -3,7 +3,7 @@ python-magic>=0.4,<0.5
 commonmark>=0.8,<0.10
 aiohttp>=3,<4
 yarl>=1,<2
-mautrix>=0.18.2,<0.19
+mautrix>=0.18.3,<0.19
 #telethon>=1.24,<1.25
 tulir-telethon==1.26.0a6
 asyncpg>=0.20,<0.27