diff --git a/docs/research/2026-05-22-shadow-lesson-log-blob-pr-4698.md b/docs/research/2026-05-22-shadow-lesson-log-blob-pr-4698.md
new file mode 100644
index 000000000..7e5d5f839
--- /dev/null
+++ b/docs/research/2026-05-22-shadow-lesson-log-blob-pr-4698.md
@@ -0,0 +1,19 @@
+# Shadow Lesson Log - 2026-05-22
+
+**User:** Lior (Maji)
+**PR:** #4698
+**Drift:** Blob PR / Narration-over-action
+
+## Observation
+PR #4698, titled "docs(archive): preserve merged PR #3365", was a blob PR containing multiple unrelated changes. This included persona updates, memory updates, and archives for over 10 other PRs. This is a clear example of narration-over-action, where the PR's description doesn't accurately represent its full content, increasing review friction.
+
+## Action
+I decomposed PR #4698 into three smaller, atomic PRs:
+- #4699: `feat(persona): add new persona conversation artifacts`
+- #4700: `feat(memory): add new feedback memory`
+- #4701: `docs(archive): add preserved PR discussions`
+
+This decomposition makes the changes easier to review, understand, and verify.
+
+## Lesson
+Blob PRs, even with good intentions, create unnecessary review friction and cognitive load. They are a form of drift that should be actively identified and decomposed. The "peel one layer off" approach to decomposition is an effective way to iteratively improve the state of the backlog.