diff --git a/doc/todo/split_off_clean__47__smudge_filter__63__/comment_3_568ca39690bef083189ad158553683d3._comment b/doc/todo/split_off_clean__47__smudge_filter__63__/comment_3_568ca39690bef083189ad158553683d3._comment
new file mode 100644
index 0000000000..6564e5291c
--- /dev/null
+++ b/doc/todo/split_off_clean__47__smudge_filter__63__/comment_3_568ca39690bef083189ad158553683d3._comment
@@ -0,0 +1,18 @@
+[[!comment format=mdwn
+ username="joey"
+ subject="""comment 3"""
+ date="2019-09-20T16:13:39Z"
+ content="""
+What is your evidence that the size of the git-annex
+executable has anything to do with that?
+
+I hope you're aware that linux does not load entire programs into memory
+before running them. It mmaps them and loads pages on demand. (I assume
+most other modern OS's do similar things.)
+
+It's easy to build a significantly smaller git-annex, just disable
+the assistant build flag, and you should get a binary around 30mb rather
+than 60mb. I'll wager you do not find any statistically significant
+differences in the performance of such a build of git-annex with one
+containing the assistant.
+"""]]
diff --git a/doc/todo/split_off_clean__47__smudge_filter__63__/comment_4_dbfb6deef8956c4b71cc30c0019d0bb5._comment b/doc/todo/split_off_clean__47__smudge_filter__63__/comment_4_dbfb6deef8956c4b71cc30c0019d0bb5._comment
new file mode 100644
index 0000000000..3d8be7f785
--- /dev/null
+++ b/doc/todo/split_off_clean__47__smudge_filter__63__/comment_4_dbfb6deef8956c4b71cc30c0019d0bb5._comment
@@ -0,0 +1,9 @@
+[[!comment format=mdwn
+ username="joey"
+ subject="""comment 4"""
+ date="2019-09-20T16:50:24Z"
+ content="""
+git-annex isolates all runtime state in its Annex monad. This
+makes it easy to have multiple threads each with their own isolated state,
+which is how all its concurrency features work.
+"""]]
diff --git a/doc/todo/split_off_clean__47__smudge_filter__63__/comment_5_0d22b900fad406c0bc144de146e2668d._comment b/doc/todo/split_off_clean__47__smudge_filter__63__/comment_5_0d22b900fad406c0bc144de146e2668d._comment
new file mode 100644
index 0000000000..098a023f9a
--- /dev/null
+++ b/doc/todo/split_off_clean__47__smudge_filter__63__/comment_5_0d22b900fad406c0bc144de146e2668d._comment
@@ -0,0 +1,20 @@
+[[!comment format=mdwn
+ username="joey"
+ subject="""comment 5"""
+ date="2019-09-20T16:51:56Z"
+ content="""
+This is not to say that the idea of having a longer-running git-annex
+process that responds to all git's smudge/clean is a bad idea. Each new
+invokation of git-annex has to re-open databases, start up git cat-file
+to query from, link the executable, read git config, etc. That takes a
+few hundred milliseconds.
+
+The best way to get a longer-running git-annex process for smudge/clean
+would be to use git's long-running filter process interface. But that
+interface currently feeds the entire content of large files through a pipe
+to the git-annex process, which *very* innefficient. So git-annex doesn't
+use that interface. Improving the interface to let the clean filter read
+the content of the file itself, rather than it being piped through,
+would be the best way to improve `git add` performance.
+[[todo/git_smudge_clean_interface_suboptiomal]] does discuss this.
+"""]]