diff --git a/changelog.d/13445.misc b/changelog.d/13445.misc
new file mode 100644
index 0000000000..17462c56f3
--- /dev/null
+++ b/changelog.d/13445.misc
@@ -0,0 +1 @@
+Add some comments about how event push actions are stored.
diff --git a/synapse/storage/databases/main/event_push_actions.py b/synapse/storage/databases/main/event_push_actions.py
index 5ddddb1cf3..5db70f9a60 100644
--- a/synapse/storage/databases/main/event_push_actions.py
+++ b/synapse/storage/databases/main/event_push_actions.py
@@ -12,6 +12,67 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
+
+"""Responsible for storing and fetching push actions / notifications.
+
+There are two main uses for push actions:
+  1. Sending out push to a user's device; and
+  2. Tracking per-room per-user notification counts (used in sync requests).
+
+For the former we simply use the `event_push_actions` table, which contains all
+the calculated actions for a given user (which were calculated by the
+`BulkPushRuleEvaluator`).
+
+For the latter we could simply count the number of rows in `event_push_actions`
+table for a given room/user, but in practice this is *very* heavyweight when
+there were a large number of notifications (due to e.g. the user never reading a
+room). Plus, keeping all push actions indefinitely uses a lot of disk space.
+
+To fix these issues, we add a new table `event_push_summary` that tracks
+per-user per-room counts of all notifications that happened before a stream
+ordering S. Thus, to get the notification count for a user / room we can simply
+query a single row in `event_push_summary` and count the number of rows in
+`event_push_actions` with a stream ordering larger than S (and as long as S is
+"recent", the number of rows needing to be scanned will be small).
+
+The `event_push_summary` table is updated via a background job that periodically
+chooses a new stream ordering S' (usually the latest stream ordering), counts
+all notifications in `event_push_actions` between the existing S and S', and
+adds them to the existing counts in `event_push_summary`.
+
+This allows us to delete old rows from `event_push_actions` once those rows have
+been counted and added to `event_push_summary` (we call this process
+"rotation").
+
+
+We need to handle when a user sends a read receipt to the room. Again this is
+done as a background process. For each receipt we clear the row in
+`event_push_summary` and count the number of notifications in
+`event_push_actions` that happened after the receipt but before S, and insert
+that count into `event_push_summary` (If the receipt happened *after* S then we
+simply clear the `event_push_summary`.)
+
+Note that its possible that if the read receipt is for an old event the relevant
+`event_push_actions` rows will have been rotated and we get the wrong count
+(it'll be too low). We accept this as a rare edge case that is unlikely to
+impact the user much (since the vast majority of read receipts will be for the
+latest event).
+
+The last complication is to handle the race where we request the notifications
+counts after a user sends a read receipt into the room, but *before* the
+background update handles the receipt (without any special handling the counts
+would be outdated). We fix this by including in `event_push_summary` the read
+receipt we used when updating `event_push_summary`, and every time we query the
+table we check if that matches the most recent read receipt in the room. If yes,
+continue as above, if not we simply query the `event_push_actions` table
+directly.
+
+Since read receipts are almost always for recent events, scanning the
+`event_push_actions` table in this case is unlikely to be a problem. Even if it
+is a problem, it is temporary until the background job handles the new read
+receipt.
+"""
+
 import logging
 from typing import TYPE_CHECKING, Dict, List, Optional, Tuple, Union, cast