diff --git a/pgxn/neon/neon_walreader.c b/pgxn/neon/neon_walreader.c
index 60eb8e1fc985..0f76514b866d 100644
--- a/pgxn/neon/neon_walreader.c
+++ b/pgxn/neon/neon_walreader.c
@@ -109,11 +109,12 @@ NeonWALReaderAllocate(int wal_segment_size, XLogRecPtr available_lsn, char *log_
 {
 	NeonWALReader *reader;
 
+	/*
+	 * Note: we allocate in TopMemoryContext, reusing the reader for all process
+	 * reads.
+	 */
 	reader = (NeonWALReader *)
-		palloc_extended(sizeof(NeonWALReader),
-						MCXT_ALLOC_NO_OOM | MCXT_ALLOC_ZERO);
-	if (!reader)
-		return NULL;
+		MemoryContextAllocZero(TopMemoryContext, sizeof(NeonWALReader));
 
 	reader->available_lsn = available_lsn;
 	reader->seg.ws_file = -1;
diff --git a/test_runner/regress/test_logical_replication.py b/test_runner/regress/test_logical_replication.py
index 41283e4d2ca0..66afe9ddfdde 100644
--- a/test_runner/regress/test_logical_replication.py
+++ b/test_runner/regress/test_logical_replication.py
@@ -247,6 +247,12 @@ def test_ondemand_wal_download_in_replication_slot_funcs(neon_env_builder: NeonE
         cur.execute(
             "SELECT * FROM pg_logical_slot_peek_binary_changes('slotty_mcslotface', NULL, NULL, 'include-xids', '0')"
         )
+        # do the peek second time: we've had a bug using wrong memory context
+        # for NeonWALReader leading to the crash in this case.
+        log.info("peek_changes again")
+        cur.execute(
+            "SELECT * FROM pg_logical_slot_peek_binary_changes('slotty_mcslotface', NULL, NULL, 'include-xids', '0')"
+        )
 
 
 # Tests that walsender correctly blocks until WAL is downloaded from safekeepers