diff --git a/.changeset/big-insert-import.md b/.changeset/big-insert-import.md
new file mode 100644
index 000000000..b568fa2d2
--- /dev/null
+++ b/.changeset/big-insert-import.md
@@ -0,0 +1,5 @@
+---
+"loro-crdt": patch
+---
+
+Fix importing one big text insert made right after a concurrent change taking quadratic time (3M characters: ~50 s, now ~0.1 s). The text tracker updated the id-to-cursor map once per rope leaf report, and the rope reported the same leaf once per 256-op fragment; each update walked the whole insert. Repeated leaves are now updated once.
diff --git a/crates/loro-internal/src/container/richtext/tracker.rs b/crates/loro-internal/src/container/richtext/tracker.rs
index 942f28730..3ff591ba4 100644
--- a/crates/loro-internal/src/container/richtext/tracker.rs
+++ b/crates/loro-internal/src/container/richtext/tracker.rs
@@ -6,6 +6,7 @@ use generic_btree::{
 };
 use loro_common::{Counter, HasId, HasIdSpan, IdFull, IdSpan, Lamport, PeerID, ID};
 use rle::HasLength as _;
+use rustc_hash::FxHashSet;
 use tracing::instrument;
 
 use crate::{cursor::AbsolutePosition, VersionVector};
@@ -186,7 +187,12 @@ impl Tracker {
             }
             _ => {
                 let mut updates = Vec::with_capacity(split.len());
+                let mut seen = FxHashSet::default();
                 for &new_leaf_idx in split {
+                    // The rope reports a leaf once per update that touched it; updating it once is enough.
+                    if !seen.insert(new_leaf_idx) {
+                        continue;
+                    }
                     let leaf = self.rope.tree().get_elem(new_leaf_idx).unwrap();
                     updates.push((leaf.id_span(), new_leaf_idx));
                 }
diff --git a/crates/loro/tests/big_insert_import.rs b/crates/loro/tests/big_insert_import.rs
new file mode 100644
index 000000000..384e86fa8
--- /dev/null
+++ b/crates/loro/tests/big_insert_import.rs
@@ -0,0 +1,67 @@
+//! Importing one big text insert made right after a concurrent change. The rope
+//! reported the inserted leaf once per id-to-cursor fragment (256 ops), and each
+//! report walked every fragment of the insert, so the import took time quadratic
+//! in the insert's length (3M characters: ~50 s).
+//!
+//! Run the timing check with `cargo test --release -p loro --test big_insert_import
+//! -- --ignored --nocapture`.
+
+use loro::{ExportMode, LoroDoc};
+use std::time::Instant;
+
+fn text(n: usize) -> String {
+    (0..n)
+        .map(|i| (b'a' + ((i as u64 * 2_654_435_761) >> 7) as u8 % 26) as char)
+        .collect()
+}
+
+/// Peer 1 inserts `n` characters at the start right after importing peer 2's
+/// concurrent edit; returns both docs and peer 1's update for the insert.
+fn big_insert(n: usize) -> (LoroDoc, LoroDoc, Vec<u8>) {
+    let a = LoroDoc::new();
+    a.set_peer_id(1).unwrap();
+    a.get_text("text").insert(0, "Hello world").unwrap();
+    a.commit();
+    let b = LoroDoc::new();
+    b.set_peer_id(2).unwrap();
+    b.import(&a.export(ExportMode::Snapshot).unwrap()).unwrap();
+    b.get_text("text").insert(0, "B ").unwrap();
+    b.commit();
+    a.import(&b.export(ExportMode::all_updates()).unwrap())
+        .unwrap();
+    let from = a.oplog_vv();
+    a.get_text("text").insert(0, &text(n)).unwrap();
+    a.commit();
+    let update = a.export(ExportMode::updates(&from)).unwrap();
+    (a, b, update)
+}
+
+#[test]
+fn big_insert_after_concurrent_edit_imports_the_same_text() {
+    let (a, b, update) = big_insert(20_000);
+    b.import(&update).unwrap();
+    assert_eq!(
+        a.get_text("text").to_string(),
+        b.get_text("text").to_string()
+    );
+    assert_eq!(
+        a.get_text("text").len_unicode(),
+        20_000 + "B Hello world".len()
+    );
+}
+
+#[test]
+#[ignore]
+fn big_insert_import_is_not_quadratic() {
+    let (a, b, update) = big_insert(1_000_000);
+    let start = Instant::now();
+    b.import(&update).unwrap();
+    let elapsed = start.elapsed().as_secs_f64();
+    println!("1M characters imported in {elapsed:.3}s");
+    assert_eq!(
+        a.get_text("text").to_string(),
+        b.get_text("text").to_string()
+    );
+    // ~0.02 s linear on a laptop; ~5 s when quadratic.
+    assert!(elapsed < 1.0, "1M-character import took {elapsed:.3}s");
+}