diff --git a/raven/importer/orchestrator.py b/raven/importer/orchestrator.py
index a38710e80..b1043da91 100644
--- a/raven/importer/orchestrator.py
+++ b/raven/importer/orchestrator.py
@@ -2,6 +2,7 @@
 
 from __future__ import annotations
 
+import asyncio
 from collections.abc import Callable, Sequence
 from dataclasses import dataclass
 from pathlib import Path
@@ -17,13 +18,27 @@
 # message_id from (session_id, timestamp_ms, index-within-batch), so those
 # boundaries are part of the id: two messages sharing a millisecond collide,
 # and one is dropped, if they land at the same index in different batches.
-# Fifty, not a hundred: EverOS extracts on every add, and that cost is
-# superlinear in the message count -- against a real service a 15-message
-# batch took 12s and a 52-message batch 24s, while a batch of 100 ran past the
-# six-minute extraction budget and failed every memory-file source.
-_BATCH_MSG_LIMIT = 50
+# Ten: EverOS extracts on every add, and that cost is superlinear in the
+# message count -- against a real service a 15-message batch took 12s and a
+# 52-message batch 24s, while a batch of 100 ran past the six-minute
+# extraction budget and failed every memory-file source. With a slower
+# extraction model, batches of 50 took 2.4-7.4 minutes and six of seven
+# memory-file sources died on that same budget. Ten is the maintainer's
+# call: a batch that finishes well inside the budget on any model matters
+# more than the fixed cost of about 7s that every add carries -- which is
+# also why it is not one message per add.
+_BATCH_MSG_LIMIT = 10
 _BATCH_CHAR_LIMIT = 30_000
 
+# A batch the memory service refuses is sent again before its source is given
+# up on. The refusals seen against a real service were transient -- a rate
+# limit at the extraction provider, an answer the extractor could not parse, a
+# slow answer past the budget -- and a source that fails on one of them takes
+# every message behind it down with it, then the next source runs straight
+# into the same wall. Three retries with these waits cover a rate-limit window
+# of a couple of minutes; the wait polls the stop file so a stop lands in it.
+_STORE_RETRY_BACKOFF_S: tuple[float, ...] = (30.0, 60.0, 120.0)
+
 
 @dataclass(frozen=True)
 class ImportFailure:
@@ -206,8 +221,10 @@ async def _feed_session(
     cancel_path: Path | None = None,
 ) -> bool:
     """Store the session in batches. Returns False when a stop request arrived
-    between two batches, leaving the rest unsent; a long conversation is many
-    batches, and a stop that waited for the whole source was not a stop."""
+    between two batches or during a retry wait, leaving the rest unsent; a long
+    conversation is many batches, and a stop that waited for the whole source
+    was not a stop. A batch the backend refuses is retried on
+    ``_STORE_RETRY_BACKOFF_S`` before the source counts as failed."""
     if not session.messages:
         return True
     all_dicts = [_to_store_dict(m) for m in session.messages]
@@ -217,36 +234,71 @@ async def _feed_session(
     def _cancelled() -> bool:
         return cancel_path is not None and cancel_path.exists()
 
-    async def _flush(*, is_final: bool) -> None:
+    async def _flush(*, is_final: bool) -> bool:
         nonlocal batch, batch_chars
         # bulk: nothing waits on an import write; the backend budgets it as extraction.
         metadata: dict[str, Any] = {"is_final": is_final, "bulk": True}
         _log_store_request(session.session_id, batch, metadata, batch_chars)
-        landed = await backend.store(session.session_id, batch, metadata=metadata)
-        if landed is False:
-            raise MemoryWriteDroppedError(
-                f"memory service did not accept a batch for {session.session_id}; source left unsubmitted"
+        for attempt, wait in enumerate((*_STORE_RETRY_BACKOFF_S, None), start=1):
+            try:
+                landed = await backend.store(session.session_id, batch, metadata=metadata)
+                reason = "memory service did not accept a batch"
+            except Exception as exc:
+                landed, reason = False, (str(exc) or repr(exc))
+            if landed is not False:
+                break
+            if wait is None:
+                raise MemoryWriteDroppedError(
+                    f"{reason} for {session.session_id} after {attempt} attempts; source left unsubmitted"
+                )
+            logger.warning(
+                "batch for {} not accepted ({}); retrying in {}s ({}/{})",
+                session.session_id,
+                reason,
+                int(wait),
+                attempt,
+                len(_STORE_RETRY_BACKOFF_S),
             )
+            if not await _pause(wait, _cancelled):
+                return False
         logger.debug("store completed: session_id={}", session.session_id)
         batch = []
         batch_chars = 0
+        return True
 
     for msg_dict in all_dicts:
         msg_chars = len(msg_dict["content"])
         if batch and (len(batch) >= _BATCH_MSG_LIMIT or batch_chars + msg_chars > _BATCH_CHAR_LIMIT):
             if _cancelled():
                 return False
-            await _flush(is_final=False)
+            if not await _flush(is_final=False):
+                return False
         batch.append(msg_dict)
         batch_chars += msg_chars
 
     if batch:
         if _cancelled():
             return False
-        await _flush(is_final=True)
+        if not await _flush(is_final=True):
+            return False
     return True
 
 
+async def _pause(seconds: float, cancelled: Callable[[], bool]) -> bool:
+    """Wait out a retry backoff a second at a time, so a stop lands inside it.
+
+    Returns False when the stop arrived."""
+    loop = asyncio.get_running_loop()
+    deadline = loop.time() + seconds
+    while True:
+        if cancelled():
+            return False
+        remaining = deadline - loop.time()
+        if remaining <= 0:
+            return True
+        await asyncio.sleep(min(1.0, remaining))
+
+
 def _log_store_request(
     session_id: str,
     batch: list[dict[str, Any]],
diff --git a/tests/integration/test_import_e2e.py b/tests/integration/test_import_e2e.py
index 601b6ccbc..a5af3595a 100644
--- a/tests/integration/test_import_e2e.py
+++ b/tests/integration/test_import_e2e.py
@@ -237,7 +237,7 @@ async def test_full_pipeline_memory_files(scanner: ClaudeCodeScanner, tmp_path:
 
 @pytest.mark.asyncio
 async def test_batching_large_conversation(scanner: ClaudeCodeScanner, tmp_path: Path) -> None:
-    """160 messages -> four store calls of 50, 50, 50 and 10, is_final only on the last, every one bulk."""
+    """160 messages -> sixteen store calls of 10, is_final only on the last, every one bulk."""
     results = await scanner.scan()
     items = _items_of_kind(scanner, results, SourceKind.CONVERSATION, source_key="sess-large")
     assert len(items) == 1
@@ -249,8 +249,8 @@ async def test_batching_large_conversation(scanner: ClaudeCodeScanner, tmp_path:
 
     assert summary.submitted == 1
     sizes = [len(call["messages"]) for call in backend.store_calls]
-    assert sizes == [50, 50, 50, 10]
-    assert [call["metadata"]["is_final"] for call in backend.store_calls] == [False, False, False, True]
+    assert sizes == [10] * 16
+    assert [call["metadata"]["is_final"] for call in backend.store_calls] == [False] * 15 + [True]
     assert all(call["metadata"]["bulk"] is True for call in backend.store_calls)
     assert sum(sizes) == 160
 
diff --git a/tests/test_importer_orchestrator.py b/tests/test_importer_orchestrator.py
index c803bd390..cdda04b70 100644
--- a/tests/test_importer_orchestrator.py
+++ b/tests/test_importer_orchestrator.py
@@ -7,6 +7,7 @@
 
 import pytest
 
+from raven.importer import orchestrator
 from raven.importer.orchestrator import ImportSummary, ProgressEvent, run_import
 from raven.importer.state import ImportState
 from raven.importer.types import (
@@ -22,15 +23,30 @@
 # ---------------------------------------------------------------------------
 
 
+@pytest.fixture(autouse=True)
+def _no_backoff(monkeypatch: pytest.MonkeyPatch) -> None:
+    """The retry waits are real seconds in production; tests that care set them."""
+    monkeypatch.setattr(orchestrator, "_STORE_RETRY_BACKOFF_S", (0.0, 0.0, 0.0))
+
+
 class FakeBackend:
     """Records store() calls for assertion."""
 
-    def __init__(self, *, fail_on: set[str] | None = None, drop_on: set[str] | None = None) -> None:
+    def __init__(
+        self,
+        *,
+        fail_on: set[str] | None = None,
+        drop_on: set[str] | None = None,
+        drop_first: dict[str, int] | None = None,
+    ) -> None:
         self.calls: list[dict[str, Any]] = []
+        self.attempts = 0
         self._fail_on = fail_on or set()
         # A backend that reports a dropped write instead of raising: the shape
         # a real EverosBackend takes when the memory service is unavailable.
         self._drop_on = drop_on or set()
+        # Refuse the first n writes of a session, then accept: a transient fault.
+        self._drop_first = dict(drop_first or {})
 
     async def recall(self, query: str, *, user_id: str | None = None, agent_id: str | None = None, top_k: int) -> list:
         return []
@@ -38,10 +54,14 @@ async def recall(self, query: str, *, user_id: str | None = None, agent_id: str
     async def store(
         self, session_id: str, messages: list[dict[str, Any]], *, metadata: dict[str, Any] | None = None
     ) -> bool:
+        self.attempts += 1
         if session_id in self._fail_on:
             raise RuntimeError(f"store failed for {session_id}")
         if session_id in self._drop_on:
             return False
+        if self._drop_first.get(session_id, 0) > 0:
+            self._drop_first[session_id] -= 1
+            return False
         self.calls.append({"session_id": session_id, "messages": messages, "metadata": metadata})
         return True
 
@@ -247,18 +267,116 @@ async def test_a_dropped_write_is_treated_exactly_like_a_raised_one(self, tmp_pa
         assert state.is_submitted("claude_code", "b")
 
 
+class TestStoreRetry:
+    """A refused batch is sent again, with a wait, before its source is given up on."""
+
+    @staticmethod
+    def _recording_pause(waits: list[float]):
+        async def _pause(seconds: float, cancelled: Any) -> bool:
+            waits.append(seconds)
+            return True
+
+        return _pause
+
+    @pytest.mark.asyncio
+    async def test_a_refused_batch_is_retried_after_each_wait_and_lands(
+        self, tmp_path: Path, monkeypatch: pytest.MonkeyPatch
+    ) -> None:
+        waits: list[float] = []
+        monkeypatch.setattr(orchestrator, "_STORE_RETRY_BACKOFF_S", (5.0, 7.0, 9.0))
+        monkeypatch.setattr(orchestrator, "_pause", self._recording_pause(waits))
+        state = ImportState(path=tmp_path / "state.json")
+        backend = FakeBackend(drop_first={"import-a": 2})
+        scanner = FakeScanner({"a": _session(n_msgs=1, session_id="import-a")})
+
+        summary = await run_import([(scanner, _scan_result("a"))], backend, state)
+
+        assert (summary.submitted, summary.failed) == (1, 0)
+        assert state.is_submitted("claude_code", "a")
+        assert backend.attempts == 3
+        assert len(backend.calls) == 1
+        assert waits == [5.0, 7.0]
+
+    @pytest.mark.asyncio
+    async def test_a_batch_refused_every_time_fails_its_source_after_the_last_wait(
+        self, tmp_path: Path, monkeypatch: pytest.MonkeyPatch
+    ) -> None:
+        waits: list[float] = []
+        monkeypatch.setattr(orchestrator, "_STORE_RETRY_BACKOFF_S", (1.0, 2.0, 3.0))
+        monkeypatch.setattr(orchestrator, "_pause", self._recording_pause(waits))
+        state = ImportState(path=tmp_path / "state.json")
+        backend = FakeBackend(drop_on={"import-a"})
+        scanner = FakeScanner({"a": _session(n_msgs=1, session_id="import-a")})
+
+        summary = await run_import([(scanner, _scan_result("a"))], backend, state)
+
+        assert (summary.submitted, summary.failed) == (0, 1)
+        assert backend.attempts == 4
+        assert waits == [1.0, 2.0, 3.0]
+        assert "after 4 attempts" in summary.errors[0].error
+
+    @pytest.mark.asyncio
+    async def test_a_raised_store_error_is_retried_the_same_way(self, tmp_path: Path) -> None:
+        state = ImportState(path=tmp_path / "state.json")
+        backend = FakeBackend(fail_on={"import-a"})
+        scanner = FakeScanner({"a": _session(n_msgs=1, session_id="import-a")})
+
+        summary = await run_import([(scanner, _scan_result("a"))], backend, state)
+
+        assert backend.attempts == 4
+        assert summary.errors[0].error.startswith("store failed for import-a for import-a after 4 attempts")
+
+    @pytest.mark.asyncio
+    async def test_a_stop_during_the_wait_ends_the_run_and_leaves_the_source_unmarked(
+        self, tmp_path: Path, monkeypatch: pytest.MonkeyPatch
+    ) -> None:
+        cancel = tmp_path / "cancel"
+
+        async def _stopped_pause(seconds: float, cancelled: Any) -> bool:
+            cancel.touch()
+            return False
+
+        monkeypatch.setattr(orchestrator, "_pause", _stopped_pause)
+        state = ImportState(path=tmp_path / "state.json")
+        backend = FakeBackend(drop_first={"import-a": 1})
+        scanner = FakeScanner(
+            {"a": _session(n_msgs=1, session_id="import-a"), "b": _session(n_msgs=1, session_id="import-b")}
+        )
+        items = [(scanner, _scan_result("a")), (scanner, _scan_result("b"))]
+
+        summary = await run_import(items, backend, state, cancel_path=cancel)
+
+        assert (summary.submitted, summary.failed) == (0, 0)
+        assert not state.is_submitted("claude_code", "a")
+        assert state.get_progress()["entries"] == {}
+        assert [c["session_id"] for c in backend.calls] == []
+
+    @pytest.mark.asyncio
+    async def test_pause_returns_early_when_the_stop_arrives(self) -> None:
+        seen = 0
+
+        def _cancelled() -> bool:
+            nonlocal seen
+            seen += 1
+            return seen > 1
+
+        assert await orchestrator._pause(0.0, lambda: False) is True
+        assert await orchestrator._pause(30.0, _cancelled) is False
+        assert seen == 2
+
+
 class TestBatching:
     @pytest.mark.asyncio
     async def test_msg_count_limit(self, tmp_path: Path) -> None:
-        """120 messages -> 3 batches (50 + 50 + 20), only the last one final, every one bulk."""
+        """120 messages -> 12 batches of 10, only the last one final, every one bulk."""
         state = ImportState(path=tmp_path / "state.json")
         backend = FakeBackend()
         scanner = FakeScanner({"k1": _session(n_msgs=120, session_id="s1", content="x")})
 
         await run_import([(scanner, _scan_result("k1"))], backend, state)
 
-        assert [len(c["messages"]) for c in backend.calls] == [50, 50, 20]
-        assert [c["metadata"]["is_final"] for c in backend.calls] == [False, False, True]
+        assert [len(c["messages"]) for c in backend.calls] == [10] * 12
+        assert [c["metadata"]["is_final"] for c in backend.calls] == [False] * 11 + [True]
         assert all(c["metadata"]["bulk"] is True for c in backend.calls)
 
     def test_a_batch_stays_inside_the_zone_everos_extracts_linearly(self) -> None:
@@ -269,7 +387,7 @@ def test_a_batch_stays_inside_the_zone_everos_extracts_linearly(self) -> None:
         memory-file source it belonged to."""
         from raven.importer.orchestrator import _BATCH_MSG_LIMIT
 
-        assert _BATCH_MSG_LIMIT <= 50
+        assert _BATCH_MSG_LIMIT <= 10
 
     @pytest.mark.asyncio
     async def test_char_limit_fallback(self, tmp_path: Path) -> None:
@@ -288,10 +406,10 @@ async def test_char_limit_fallback(self, tmp_path: Path) -> None:
 
     @pytest.mark.asyncio
     async def test_is_final_only_on_last_batch(self, tmp_path: Path) -> None:
-        """Exactly 50 messages -> 1 batch with is_final=True."""
+        """Exactly 10 messages -> 1 batch with is_final=True."""
         state = ImportState(path=tmp_path / "state.json")
         backend = FakeBackend()
-        scanner = FakeScanner({"k1": _session(n_msgs=50, session_id="s1", content="x")})
+        scanner = FakeScanner({"k1": _session(n_msgs=10, session_id="s1", content="x")})
 
         await run_import([(scanner, _scan_result("k1"))], backend, state)
 
diff --git a/ui-web/src/features/importSync/ImportSyncPage.test.tsx b/ui-web/src/features/importSync/ImportSyncPage.test.tsx
index 437e8f1ec..19008d26d 100644
--- a/ui-web/src/features/importSync/ImportSyncPage.test.tsx
+++ b/ui-web/src/features/importSync/ImportSyncPage.test.tsx
@@ -84,14 +84,18 @@ describe('the import row', () => {
     expect(runs).toEqual([[['claude_code'], 'memory_files']])
   })
 
-  it('says how many did not make it and offers a retry, with nothing to dismiss', () => {
+  it('says how many did not make it, offers a retry, and a dismiss that takes the row down', async () => {
     draw(status({ total: 18, submitted: 15, failed: 3, tier: 'full', platforms: ['claude_code'], phases: { status: 'done', errors: [] } }))
     expect(row()?.className).toContain('importSync-done')
     expect(row()?.className).toContain('importSync-warn')
     expect(row()?.textContent).toContain('gui.importSync.failed_n:{"n":3}')
     expect(main().disabled).toBe(false)
     expect(main().getAttribute('aria-label')).toBe('gui.importSync.retry')
-    expect(x()).toBeNull()
+    expect(x()?.getAttribute('aria-label')).toBe('gui.importSync.dismiss')
+
+    await act(async () => { fireEvent.click(x()!) })
+
+    expect(row()).toBeNull()
   })
 
   it('shows a stopped run it cannot ask for again as paused with nothing to click', () => {
diff --git a/ui-web/src/features/importSync/ImportSyncPage.tsx b/ui-web/src/features/importSync/ImportSyncPage.tsx
index b59b462bb..a1b37285d 100644
--- a/ui-web/src/features/importSync/ImportSyncPage.tsx
+++ b/ui-web/src/features/importSync/ImportSyncPage.tsx
@@ -61,7 +61,7 @@ export function ImportSyncApp(): JSX.Element | null {
         </button>
         {running ? (
           <button type="button" className="importSync-x" aria-label={t('gui.importSync.stop')} onClick={() => void store.stop()}><Close /></button>
-        ) : v.kind === 'done' && !v.failed ? (
+        ) : v.kind === 'done' ? (
           <button type="button" className="importSync-x" aria-label={t('gui.importSync.dismiss')} onClick={store.dismiss}><Close /></button>
         ) : null}
       </div>