diff --git a/doc/impl-thoughts/REPLIES.md b/doc/impl-thoughts/REPLIES.md
new file mode 100644
index 00000000..eac5b7ea
--- /dev/null
+++ b/doc/impl-thoughts/REPLIES.md
@@ -0,0 +1,17 @@
+If we were to render replies in a smart way (instead of relying on the fallback), we would
+need to manually find entries that are pointed to be `in_reply_to`. Consulting the timeline
+code, it seems appropriate to add a `_replyingTo` field to a `BaseEventEntry` (much like we
+have `_pendingAnnotations` and `pendingRedactions`). We can then:
+* use `TilesCollection`'s `_findTileIdx` to find the tile of the message being replied to,
+  and put a reference to its tile into the new tile being created (?).
+  * It doesn't seem appropriate to add an additional argument to TileCreator, but we may
+    want to re-use tiles instead of creating duplicate ones. Otherwise, of course, `tileCreator`
+    can create more than one tile from an entry's `_replyingTo` field.
+* Resolve `_replyingTo` much like we resolve `redactingEntry` in timeline: search by `relatedTxnId`
+  and `relatedEventId` if our entry is a reply (we can add an `isReply` flag there).
+  * This works fine for local entries, which are loaded via an `AsyncMappedList`, but what
+    about remote entries? They are not loaded asynchronously, and the fact that they are
+    not a derived collection is used throughout `Timeline`.
+* Entries that don't have replies that are loadeded (but that are replies) probably need
+  to be tracked somehow?
+  * Then, on timeline add, check new IDs and update corresponding entries