diff --git a/changelog.d/12179.doc b/changelog.d/12179.doc
new file mode 100644
index 000000000..55d8caa45
--- /dev/null
+++ b/changelog.d/12179.doc
@@ -0,0 +1 @@
+Updates to the Room DAG concepts development document.
diff --git a/docs/development/room-dag-concepts.md b/docs/development/room-dag-concepts.md
index cbc7cf294..3eb4d5acc 100644
--- a/docs/development/room-dag-concepts.md
+++ b/docs/development/room-dag-concepts.md
@@ -30,13 +30,57 @@ rather than skipping any that arrived late; whereas if you're looking at a
 historical section of timeline (i.e. `/messages`), you want to see the best
 representation of the state of the room as others were seeing it at the time.
 
+## Outliers
+
+We mark an event as an `outlier` when we haven't figured out the state for the
+room at that point in the DAG yet. They are "floating" events that we haven't
+yet correlated to the DAG.
+
+Outliers typically arise when we fetch the auth chain or state for a given
+event. When that happens, we just grab the events in the state/auth chain,
+without calculating the state at those events, or backfilling their
+`prev_events`.
+
+So, typically, we won't have the `prev_events` of an `outlier` in the database,
+(though it's entirely possible that we *might* have them for some other
+reason). Other things that make outliers different from regular events:
+
+ * We don't have state for them, so there should be no entry in
+   `event_to_state_groups` for an outlier. (In practice this isn't always
+   the case, though I'm not sure why: see https://github.com/matrix-org/synapse/issues/12201).
+
+ * We don't record entries for them in the `event_edges`,
+   `event_forward_extremeties` or `event_backward_extremities` tables.
+
+Since outliers are not tied into the DAG, they do not normally form part of the
+timeline sent down to clients via `/sync` or `/messages`; however there is an
+exception:
+
+### Out-of-band membership events
+
+A special case of outlier events are some membership events for federated rooms
+that we aren't full members of. For example:
+
+ * invites received over federation, before we join the room
+ * *rejections* for said invites
+ * knock events for rooms that we would like to join but have not yet joined.
+
+In all the above cases, we don't have the state for the room, which is why they
+are treated as outliers. They are a bit special though, in that they are
+proactively sent to clients via `/sync`.
 
 ## Forward extremity
 
-Most-recent-in-time events in the DAG which are not referenced by any other events' `prev_events` yet.
+Most-recent-in-time events in the DAG which are not referenced by any other
+events' `prev_events` yet. (In this definition, outliers, rejected events, and
+soft-failed events don't count.)
 
-The forward extremities of a room are used as the `prev_events` when the next event is sent.
+The forward extremities of a room (or at least, a subset of them, if there are
+more than ten) are used as the `prev_events` when the next event is sent.
 
+The "current state" of a room (ie: the state which would be used if we
+generated a new event) is, therefore, the resolution of the room states
+at each of the forward extremities.
 
 ## Backward extremity
 
@@ -44,23 +88,14 @@ The current marker of where we have backfilled up to and will generally be the
 `prev_events` of the oldest-in-time events we have in the DAG. This gives a starting point when
 backfilling history.
 
-When we persist a non-outlier event, we clear it as a backward extremity and set
-all of its `prev_events` as the new backward extremities if they aren't already
-persisted in the `events` table.
-
-
-## Outliers
-
-We mark an event as an `outlier` when we haven't figured out the state for the
-room at that point in the DAG yet.
-
-We won't *necessarily* have the `prev_events` of an `outlier` in the database,
-but it's entirely possible that we *might*.
-
-For example, when we fetch the event auth chain or state for a given event, we
-mark all of those claimed auth events as outliers because we haven't done the
-state calculation ourself.
+Note that, unlike forward extremities, we typically don't have any backward
+extremity events themselves in the database - or, if we do, they will be "outliers" (see
+above). Either way, we don't expect to have the room state at a backward extremity.
 
+When we persist a non-outlier event, if it was previously a backward extremity,
+we clear it as a backward extremity and set all of its `prev_events` as the new
+backward extremities if they aren't already persisted as non-outliers. This
+therefore keeps the backward extremities up-to-date.
 
 ## State groups