diff --git a/backend/app/context/__init__.py b/backend/app/context/__init__.py
index fef581e..3cfd2f2 100644
--- a/backend/app/context/__init__.py
+++ b/backend/app/context/__init__.py
@@ -1,3 +1,3 @@
-from .builder import build_context, count_tokens, truncate_to_last_tokens
+from .builder import build_context, count_tokens, story_actions, truncate_to_last_tokens
 
-__all__ = ["build_context", "count_tokens", "truncate_to_last_tokens"]
+__all__ = ["build_context", "count_tokens", "story_actions", "truncate_to_last_tokens"]
diff --git a/backend/app/context/builder.py b/backend/app/context/builder.py
index f84af84..2d69cf3 100644
--- a/backend/app/context/builder.py
+++ b/backend/app/context/builder.py
@@ -74,10 +74,24 @@ def _history_text(action: models.Action) -> str:
     return text
 
 
-def _visible_npcs(adventure: models.Adventure, stat_schema: dict) -> dict[str, str]:
+def story_actions(
+    adventure: models.Adventure, exclude_action_id: int | None = None
+) -> list[models.Action]:
+    """The adventure's non-empty actions, as the model should see them.
+
+    `exclude_action_id` drops one action from the story — used by retry, where
+    the row being regenerated is still attached to the adventure (it holds the
+    variant history) but must not appear in the context assembled to replace it.
+    """
+    return [
+        a for a in adventure.actions
+        if a.text.strip() and (exclude_action_id is None or a.id != exclude_action_id)
+    ]
+
+
+def _visible_npcs(actions: list[models.Action], stat_schema: dict) -> dict[str, str]:
     """Defined NPCs whose trigger words appear in the recent story — the ones
     "in scene", so only their stats get injected. Maps npc id -> display name."""
-    actions = [a for a in adventure.actions if a.text.strip()]
     recent = SEPARATOR.join(a.text for a in actions[-6:]).lower()
     visible: dict[str, str] = {}
     for npc_key, ndef in (stat_schema.get("npcs") or {}).items():
@@ -107,10 +121,13 @@ def build_context(
     adventure: models.Adventure,
     settings: models.Settings,
     memory_bank: dict | None = None,
+    exclude_action_id: int | None = None,
 ) -> tuple[str, str, dict]:
     """Returns (system_text, story_text, context_report). `memory_bank` is the
-    result of memorybank.retrieve_memories (None when the bank is off)."""
+    result of memorybank.retrieve_memories (None when the bank is off);
+    `exclude_action_id` omits one action from the story (see story_actions)."""
     script_mem = _script_memory(adventure)
+    actions = story_actions(adventure, exclude_action_id)
 
     # ----- Always-included components -----
     system_sections: list[Section] = [Section("narrator", settings.narrator_prompt.strip())]
@@ -123,7 +140,7 @@ def build_context(
         if guide:
             system_sections.append(Section("world_state_guide", guide))
         block = worldstate.render_state_section(
-            adventure.world_state, stat_schema, _visible_npcs(adventure, stat_schema)
+            adventure.world_state, stat_schema, _visible_npcs(actions, stat_schema)
         )
         if block:
             system_sections.append(Section("world_state", block))
@@ -163,7 +180,6 @@ def build_context(
     available = max(256, settings.context_token_budget - reserved)
 
     # ----- Story cards: triggered by recent story text (the window history could fill) -----
-    actions = [a for a in adventure.actions if a.text.strip()]
     trigger_window = truncate_to_last_tokens(SEPARATOR.join(a.text for a in actions), available)
     triggered = _match_cards(adventure.story_cards, trigger_window)
 
diff --git a/backend/app/memorybank.py b/backend/app/memorybank.py
index dbb785e..ad9f81a 100644
--- a/backend/app/memorybank.py
+++ b/backend/app/memorybank.py
@@ -4,7 +4,9 @@
 After each turn, a fire-and-forget task (`run_post_turn`) runs with its own DB
 session:
   - every MEMORY_INTERVAL actions (starting at MEMORY_START), each uncovered
-    block of actions is summarized into a short "memory";
+    block of actions is summarized into a short "memory". Summarization only
+    ever reads *settled* actions (see settled_story_actions) — the newest action
+    is held back one turn because it is still retryable;
   - every SUMMARY_INTERVAL actions, the Story Summary is rewritten folding in
     the new memories (the user-edited text is always the base, never clobbered);
   - new memories are embedded (OpenAI-compatible /v1/embeddings) and the bank
@@ -23,7 +25,7 @@ import math
 from sqlalchemy.orm import Session
 
 from . import models
-from .context import truncate_to_last_tokens
+from .context import story_actions, truncate_to_last_tokens
 from .database import SessionLocal
 from .providers import OpenAICompatibleProvider, ProviderError
 
@@ -86,22 +88,78 @@ def cosine(a: list[float], b: list[float]) -> float:
     return dot / norm if norm else 0.0
 
 
-def story_actions(adventure: models.Adventure) -> list[models.Action]:
-    return [a for a in adventure.actions if a.text.strip()]
+def settled_story_actions(adventure: models.Adventure) -> list[models.Action]:
+    """Story actions old enough to summarize: everything but the newest one.
+
+    Only the *last* action can be retried, so once an action has another action
+    after it, its text is final. Summarizing right up to the newest action meant
+    a memory could describe an attempt the player then retried away — the
+    memory's cursor has already advanced, so it is never regenerated, leaving a
+    memory (and, downstream, a story summary) describing narration that is no
+    longer in the story. Holding one action back costs a turn of latency and
+    makes that unreachable.
+
+    The result is always a prefix of story_actions(), so memory_cursor and
+    summary_cursor stay valid positions and no action is ever skipped.
+    """
+    return story_actions(adventure)[:-1]
+
+
+def _rewind_cursors_to_index(adventure: models.Adventure, index: int) -> None:
+    """Move both cursors back to the position of Action.index `index`.
+
+    The cursors are *positions* into story_actions() while Memory.source_* are
+    Action.index values, so the two spaces have to be translated between (they
+    diverge as soon as any action is deleted).
+    """
+    actions = story_actions(adventure)
+    position = next((i for i, a in enumerate(actions) if a.index >= index), len(actions))
+    adventure.memory_cursor = min(adventure.memory_cursor, position)
+    adventure.summary_cursor = min(adventure.summary_cursor, position)
+
+
+def note_action_removed(adventure: models.Adventure, action: models.Action) -> None:
+    """Keep the cursors pointing at the same actions when one is deleted from
+    *before* them. Call BEFORE the delete, while the action is still in the list.
+
+    memory_cursor counts actions from the start of the story, so removing an
+    earlier action slides every later one down a slot — without this, an action
+    that was never summarized shifts into the "already covered" range and is
+    skipped forever.
+    """
+    actions = story_actions(adventure)
+    position = next((i for i, a in enumerate(actions) if a.id == action.id), None)
+    if position is None:
+        return
+    if position < adventure.memory_cursor:
+        adventure.memory_cursor -= 1
+    if position < adventure.summary_cursor:
+        adventure.summary_cursor -= 1
 
 
 def prune_dangling_memories(adventure: models.Adventure, db: Session) -> int:
     """Delete memories that summarized actions which no longer exist (e.g. after
     undo). source_start/source_end are Action.index values; a memory is dangling
     if any covered action is past the current end of the story. Returns the count
-    removed. Cursors are self-healing in run_post_turn, so this is cleanup only."""
+    removed.
+
+    Throwing a memory away is not enough on its own: the actions it covered are
+    still behind memory_cursor, so they would read as summarized with nothing
+    describing them. Rewind to where the earliest discarded memory began, so
+    those actions are summarized again.
+    """
     max_index = max((a.index for a in adventure.actions), default=-1)
     dangling = [
         m for m in adventure.memories
         if m.source_end is not None and m.source_end > max_index
     ]
+    if not dangling:
+        return 0
+    starts = [m.source_start for m in dangling if m.source_start is not None]
     for m in dangling:
         db.delete(m)
+    if starts:
+        _rewind_cursors_to_index(adventure, min(starts))
     return len(dangling)
 
 
@@ -112,11 +170,13 @@ async def retrieve_memories(
     settings: models.Settings,
     *,
     update_stats: bool,
+    exclude_action_id: int | None = None,
 ) -> dict | None:
     """Returns {"used": [{id, text, similarity, pinned}], "error": str|None},
     or None when the memory bank is off for this adventure. `update_stats`
-    bumps use counters (real turns only, not Insights dry runs); the caller's
-    commit persists them."""
+    bumps use counters (real turns only, not Insights dry runs).
+    `exclude_action_id` drops the action being retried from the similarity
+    query, so the discarded attempt can't steer which memories come back."""
     if not adventure.memory_bank_enabled:
         return None
     if not settings.embedding_model.strip():
@@ -126,7 +186,7 @@ async def retrieve_memories(
     if not candidates:
         return {"used": [], "error": None}
 
-    actions = story_actions(adventure)
+    actions = story_actions(adventure, exclude_action_id)
     query = truncate_to_last_tokens(
         "\n\n".join(a.text for a in actions[-4:]), RETRIEVAL_WINDOW_TOKENS
     )
@@ -198,6 +258,12 @@ async def run_post_turn(adventure_id: int) -> None:
             return
         # Undo/retry can shrink the action list below a stored cursor, which
         # would stall summarization until the story grew past it again.
+        # Deliberately the FULL count, not the settled one: an adventure that
+        # was caught up under the old rule can have a cursor equal to the action
+        # count, and clamping to settled would rewind it one step, re-covering
+        # an already-summarized action in the next block. Both consumers below
+        # read settled actions and bail on a negative remainder, so a cursor
+        # briefly sitting one past the settled end is harmless.
         count = len(story_actions(adventure))
         adventure.memory_cursor = min(adventure.memory_cursor, count)
         adventure.summary_cursor = min(adventure.summary_cursor, count)
@@ -215,7 +281,7 @@ async def run_post_turn(adventure_id: int) -> None:
 async def _create_due_memories(
     adventure: models.Adventure, settings: models.Settings, db: Session
 ) -> None:
-    actions = story_actions(adventure)
+    actions = settled_story_actions(adventure)
     provider = summary_provider(settings)
     for _ in range(MAX_MEMORIES_PER_RUN):
         cursor = adventure.memory_cursor
@@ -246,7 +312,7 @@ async def _create_due_memories(
 async def _update_story_summary(
     adventure: models.Adventure, settings: models.Settings, db: Session
 ) -> None:
-    actions = story_actions(adventure)
+    actions = settled_story_actions(adventure)
     if len(actions) - adventure.summary_cursor < SUMMARY_INTERVAL:
         return
 
diff --git a/backend/app/routers/adventures.py b/backend/app/routers/adventures.py
index 6e96113..88df0f0 100644
--- a/backend/app/routers/adventures.py
+++ b/backend/app/routers/adventures.py
@@ -506,6 +506,11 @@ async def _generate_turn(
 ):
     settings = get_settings(db, user)
     cfg = auth.resolve_provider_config(settings)
+    # On a retry the row being regenerated is still attached to the adventure
+    # (it carries the variant history), so it has to be filtered out of the
+    # context — otherwise the model is shown the attempt it is replacing as if
+    # it were established story, and writes a continuation of it.
+    replacing_id = retry_of.id if retry_of is not None else None
     if cfg.using_demo:
         # No embedding/summarization calls on the server-funded key: memory
         # retrieval is skipped (with a visible note when the bank is on).
@@ -515,12 +520,16 @@ async def _generate_turn(
             else None
         )
     else:
-        memories = await memorybank.retrieve_memories(adventure, settings, update_stats=True)
+        memories = await memorybank.retrieve_memories(
+            adventure, settings, update_stats=True, exclude_action_id=replacing_id
+        )
     # Scoreboard as it stands before this AI turn's context/output hooks mutate
     # it — stapled onto the AI action so retry can start over from here.
     state_before = snapshot_state(adventure)
     world_state_before = snapshot_world_state(adventure)
-    system_text, story_text, snapshot = build_context(adventure, settings, memories)
+    system_text, story_text, snapshot = build_context(
+        adventure, settings, memories, exclude_action_id=replacing_id
+    )
 
     # onModelContext: scripts see (and may rewrite) the whole assembled context.
     combined = f"{system_text}\n\n{story_text}" if system_text else story_text
@@ -863,9 +872,11 @@ def undo_turn(
         last = actions.pop()
         # The earliest action removed in this turn holds the pre-turn scoreboard.
         first_removed = last
+        memorybank.note_action_removed(adventure, last)
         db.delete(last)
         if last.type == "ai" and actions and actions[-1].type in ("do", "say", "story"):
             first_removed = actions.pop()
+            memorybank.note_action_removed(adventure, first_removed)
             db.delete(first_removed)
         if first_removed.state_before is not None:
             adventure.script_state = copy.deepcopy(first_removed.state_before)
@@ -1480,9 +1491,15 @@ def delete_action(
     db: Session = Depends(get_db),
     user: models.User = CurrentUser,
 ):
-    get_adventure_or_404(adventure_id, db, user)
+    adventure = get_adventure_or_404(adventure_id, db, user)
     action = db.get(models.Action, action_id)
     if action is None or action.adventure_id != adventure_id:
         raise HTTPException(404, "Action not found")
+    # Cursor bookkeeping, same as undo: slide the cursors down if this action
+    # sits before them, then drop any memory left describing a deleted action.
+    memorybank.note_action_removed(adventure, action)
     db.delete(action)
+    db.flush()  # apply the delete so pruning sees the shrunken action list
+    db.expire(adventure, ["actions"])
+    memorybank.prune_dangling_memories(adventure, db)
     db.commit()
diff --git a/backend/tests/test_memory_settling.py b/backend/tests/test_memory_settling.py
new file mode 100644
index 0000000..26cf80c
--- /dev/null
+++ b/backend/tests/test_memory_settling.py
@@ -0,0 +1,283 @@
+"""Memories must never describe an attempt the player can still retry away.
+
+Only the last action is retryable, so summarization holds the newest action
+back one turn (memorybank.settled_story_actions). Without that, a memory could
+cover the just-generated AI turn; retrying it rewrites Action.text but the
+memory cursor has already advanced, so the memory is never regenerated and goes
+on describing narration that is no longer in the story.
+
+    python -m pytest tests/test_memory_settling.py -v
+"""
+import asyncio
+import os
+import tempfile
+
+_tmp = tempfile.NamedTemporaryFile(suffix=".db", delete=False)
+_tmp.close()
+os.environ["AIDND_DB_PATH"] = _tmp.name
+os.environ.pop("AIDND_DATABASE_URL", None)
+os.environ.pop("DATABASE_URL", None)
+
+import pytest
+
+from app import memorybank, models
+from app.database import Base, SessionLocal, engine
+
+
+class StubSummarizer:
+    """Records every excerpt handed to the summarizer."""
+
+    def __init__(self):
+        self.excerpts: list[str] = []
+
+    async def complete(self, system, user, **kwargs):
+        self.excerpts.append(user)
+        return f"Memory {len(self.excerpts)}."
+
+
+@pytest.fixture()
+def db():
+    Base.metadata.create_all(bind=engine)
+    session = SessionLocal()
+    try:
+        yield session
+    finally:
+        session.close()
+        Base.metadata.drop_all(bind=engine)
+
+
+def make_adventure(db, action_count: int) -> models.Adventure:
+    """An adventure whose actions alternate player/AI, newest last."""
+    user = models.User(is_guest=False, email="memory@example.com")
+    db.add(user)
+    db.flush()
+    settings = models.Settings(user_id=user.id, api_key="enc:dummy", model="test-model")
+    db.add(settings)
+    adventure = models.Adventure(
+        user_id=user.id, title="Cave", script_state={}, auto_summarize=True
+    )
+    db.add(adventure)
+    db.flush()
+    for i in range(action_count):
+        db.add(models.Action(
+            adventure_id=adventure.id, index=i,
+            type="ai" if i % 2 else "do", text=f"Action {i}.",
+        ))
+    db.commit()
+    db.refresh(adventure)
+    return adventure
+
+
+def run_memories(db, adventure, stub, monkeypatch):
+    monkeypatch.setattr(memorybank, "summary_provider", lambda s: stub)
+    settings = db.query(models.Settings).first()
+    asyncio.run(memorybank._create_due_memories(adventure, settings, db))
+
+
+# ------------------------------------------------------------------- settling
+
+def test_settled_actions_holds_back_the_newest(db):
+    adventure = make_adventure(db, 5)
+    settled = memorybank.settled_story_actions(adventure)
+    assert [a.index for a in settled] == [0, 1, 2, 3]
+
+
+def test_settled_actions_is_a_prefix_so_cursors_stay_valid(db):
+    """The safety property behind the whole approach: dropping the newest
+    action can never renumber or skip an earlier one."""
+    adventure = make_adventure(db, 9)
+    full = memorybank.story_actions(adventure)
+    settled = memorybank.settled_story_actions(adventure)
+    assert full[: len(settled)] == settled
+
+
+def test_settled_actions_on_a_one_action_story(db):
+    adventure = make_adventure(db, 1)
+    assert memorybank.settled_story_actions(adventure) == []
+
+
+# ------------------------------------------------------- the bug this prevents
+
+def test_memory_never_covers_the_newest_retryable_action(db, monkeypatch):
+    """cursor=6 with 12 actions is exactly the case that used to bite: the
+    6-action block ends on the newest action, which is still retryable."""
+    adventure = make_adventure(db, 12)
+    adventure.memory_cursor = 6
+    db.commit()
+
+    stub = StubSummarizer()
+    run_memories(db, adventure, stub, monkeypatch)
+
+    assert stub.excerpts == []  # only 11 settled — one short of a block
+    assert db.query(models.Memory).count() == 0
+    assert adventure.memory_cursor == 6
+
+
+def test_the_block_lands_a_turn_later_without_the_newest_action(db, monkeypatch):
+    """One more action and the same block is summarized — minus the new one."""
+    adventure = make_adventure(db, 13)
+    adventure.memory_cursor = 6
+    db.commit()
+
+    stub = StubSummarizer()
+    run_memories(db, adventure, stub, monkeypatch)
+
+    assert len(stub.excerpts) == 1
+    excerpt = stub.excerpts[0]
+    assert "Action 11." in excerpt  # the block's real last action
+    assert "Action 12." not in excerpt  # the newest, still retryable
+    memory = db.query(models.Memory).one()
+    assert (memory.source_start, memory.source_end) == (6, 11)
+    assert adventure.memory_cursor == 12
+
+
+def test_first_memory_waits_one_action_past_memory_start(db, monkeypatch):
+    adventure = make_adventure(db, memorybank.MEMORY_START)
+    stub = StubSummarizer()
+    run_memories(db, adventure, stub, monkeypatch)
+    assert stub.excerpts == []
+
+    db.add(models.Action(
+        adventure_id=adventure.id, index=memorybank.MEMORY_START, type="do", text="Later.",
+    ))
+    db.commit()
+    db.refresh(adventure)
+    run_memories(db, adventure, stub, monkeypatch)
+    # 12 settled actions = two full blocks, caught up in one run (MAX_MEMORIES_
+    # PER_RUN allows 5); neither may reach the newly added newest action.
+    assert len(stub.excerpts) == 2
+    assert not any("Later." in e for e in stub.excerpts)
+
+
+def test_legacy_caught_up_adventure_is_not_rewound(db, monkeypatch):
+    """An adventure summarized under the OLD rule can have memory_cursor equal
+    to its action count. The run_post_turn clamp must use the FULL count, not
+    the settled one — clamping to settled would rewind the cursor a step and
+    re-cover an already-summarized action in the next block."""
+    adventure = make_adventure(db, 12)
+    db.add(models.Memory(adventure_id=adventure.id, text="A", source_start=0, source_end=5))
+    db.add(models.Memory(adventure_id=adventure.id, text="B", source_start=6, source_end=11))
+    adventure.memory_cursor = 12
+    adventure.summary_cursor = 12
+    db.commit()
+
+    # The clamp as run_post_turn applies it.
+    count = len(memorybank.story_actions(adventure))
+    adventure.memory_cursor = min(adventure.memory_cursor, count)
+    assert adventure.memory_cursor == 12  # not rewound to 11
+
+    # Grow the story and let the next block form.
+    for i in range(12, 25):
+        db.add(models.Action(adventure_id=adventure.id, index=i, type="do", text=f"Action {i}."))
+    db.commit()
+    db.refresh(adventure)
+    run_memories(db, adventure, StubSummarizer(), monkeypatch)
+
+    covered: dict[int, int] = {}
+    for m in db.query(models.Memory).all():
+        for i in range(m.source_start, m.source_end + 1):
+            covered[i] = covered.get(i, 0) + 1
+    assert [i for i, c in covered.items() if c > 1] == []  # no action summarized twice
+    assert [i for i in range(max(covered) + 1) if i not in covered] == []  # no gaps
+
+
+def test_no_memories_before_memory_start(db, monkeypatch):
+    adventure = make_adventure(db, 8)
+    stub = StubSummarizer()
+    run_memories(db, adventure, stub, monkeypatch)
+    assert stub.excerpts == []
+
+
+# ------------------------------------------- deleting already-summarized ground
+
+def orphans(db, adventure) -> list[int]:
+    """Action indices the cursor calls summarized that no memory describes."""
+    covered: set[int] = set()
+    for m in db.query(models.Memory).filter_by(adventure_id=adventure.id):
+        covered |= set(range(m.source_start, m.source_end + 1))
+    actions = memorybank.story_actions(adventure)
+    return [a.index for a in actions[: adventure.memory_cursor] if a.index not in covered]
+
+
+def summarized_adventure(db):
+    """13 actions with two memories covering indices 0-11, cursor at 12."""
+    adventure = make_adventure(db, 13)
+    db.add(models.Memory(adventure_id=adventure.id, text="A", source_start=0, source_end=5))
+    db.add(models.Memory(adventure_id=adventure.id, text="B", source_start=6, source_end=11))
+    adventure.memory_cursor = 12
+    adventure.summary_cursor = 12
+    db.commit()
+    db.refresh(adventure)
+    return adventure
+
+
+def test_deleting_a_middle_action_does_not_skip_a_later_one(db):
+    """memory_cursor counts positions, so removing an earlier action slides a
+    never-summarized one into the covered range unless the cursor slides too."""
+    adventure = summarized_adventure(db)
+    victim = db.query(models.Action).filter_by(adventure_id=adventure.id, index=5).one()
+
+    memorybank.note_action_removed(adventure, victim)
+    db.delete(victim)
+    db.commit()
+    db.refresh(adventure)
+
+    assert adventure.memory_cursor == 11  # slid down by one
+    assert orphans(db, adventure) == []
+
+
+def test_deleting_a_later_action_leaves_cursors_alone(db):
+    """Only actions *before* the cursor shift it."""
+    adventure = summarized_adventure(db)
+    victim = db.query(models.Action).filter_by(adventure_id=adventure.id, index=12).one()
+
+    memorybank.note_action_removed(adventure, victim)
+    db.delete(victim)
+    db.commit()
+    db.refresh(adventure)
+
+    assert adventure.memory_cursor == 12
+    assert orphans(db, adventure) == []
+
+
+def test_pruning_a_memory_rewinds_to_where_it_started(db):
+    """Discarding a memory isn't enough — the actions it covered are still
+    behind the cursor, so they must be handed back to the summarizer."""
+    adventure = summarized_adventure(db)
+    # Delete back past index 11, so memory B (6..11) covers a missing action.
+    for index in (12, 11):
+        victim = db.query(models.Action).filter_by(adventure_id=adventure.id, index=index).one()
+        memorybank.note_action_removed(adventure, victim)
+        db.delete(victim)
+    db.flush()
+    db.expire(adventure, ["actions"])
+
+    assert memorybank.prune_dangling_memories(adventure, db) == 1
+    db.commit()
+    db.refresh(adventure)
+
+    assert [m.text for m in db.query(models.Memory).all()] == ["A"]
+    assert adventure.memory_cursor == 6  # back to where the discarded memory began
+    assert orphans(db, adventure) == []
+
+
+def test_repeated_deletes_never_orphan_an_action(db):
+    """The scenario that motivated this: undo/delete-last, over and over."""
+    adventure = summarized_adventure(db)
+    for _ in range(6):
+        actions = memorybank.story_actions(adventure)
+        if not actions:
+            break
+        victim = max(actions, key=lambda a: a.index)
+        memorybank.note_action_removed(adventure, victim)
+        db.delete(victim)
+        db.flush()
+        db.expire(adventure, ["actions"])
+        memorybank.prune_dangling_memories(adventure, db)
+        count = len(memorybank.story_actions(adventure))
+        adventure.memory_cursor = min(adventure.memory_cursor, count)
+        adventure.summary_cursor = min(adventure.summary_cursor, count)
+        db.commit()
+        db.refresh(adventure)
+        assert orphans(db, adventure) == []
+        assert adventure.memory_cursor <= len(memorybank.story_actions(adventure))
diff --git a/backend/tests/test_retry_variants.py b/backend/tests/test_retry_variants.py
index 032d14e..adfcf70 100644
--- a/backend/tests/test_retry_variants.py
+++ b/backend/tests/test_retry_variants.py
@@ -39,6 +39,7 @@ class ScriptedProvider:
     """Streams the next canned reply each call, so successive retries differ."""
     replies: list = []
     calls = 0
+    prompts: list = []  # every assembled (system, story) pair, for context assertions
 
     def __init__(self, *a, **k):
         pass
@@ -46,6 +47,7 @@ class ScriptedProvider:
     async def generate(self, parts: PromptParts, *, temperature, max_tokens):
         index = min(ScriptedProvider.calls, len(ScriptedProvider.replies) - 1)
         ScriptedProvider.calls += 1
+        ScriptedProvider.prompts.append((parts.system, parts.story))
         reply = ScriptedProvider.replies[index]
         if isinstance(reply, Exception):
             raise reply
@@ -79,6 +81,7 @@ def client(monkeypatch):
 
     ScriptedProvider.replies = ["Attempt one."]
     ScriptedProvider.calls = 0
+    ScriptedProvider.prompts = []
     monkeypatch.setattr(adventures, "OpenAICompatibleProvider", ScriptedProvider)
     monkeypatch.setattr(auth, "resolve_provider_config", lambda s: auth.ProviderConfig(
         "http://fake", "k", "test-model", False))
@@ -146,6 +149,35 @@ def test_retry_keeps_the_discarded_attempt(client):
     assert [v["active"] for v in r.json()] == [False, True]
 
 
+def test_retry_context_excludes_the_attempt_being_replaced(client):
+    """The whole point of a retry is a fresh take on the *same* turn. The row
+    now survives the retry (it holds the variant history), so it is still in
+    `adventure.actions` while the replacement context is assembled — it must be
+    filtered out, or the model is asked to continue *past* the attempt it is
+    supposed to be replacing and writes a sequel that blends both."""
+    ScriptedProvider.replies = ["Attempt one.", "Attempt two."]
+    _play(client)
+    _retry(client)
+
+    retry_story = ScriptedProvider.prompts[-1][1]
+    assert "Attempt one." not in retry_story
+    # The turn's own player action must still be there — it's what to respond to.
+    assert "look around" in retry_story
+    assert "You enter a cave." in retry_story
+
+
+def test_retry_context_keeps_earlier_ai_turns(client):
+    """Only the action being retried is dropped, not AI history in general."""
+    ScriptedProvider.replies = ["First turn.", "Second turn.", "Second, again."]
+    _play(client, "go north")
+    _play(client, "go south")
+    _retry(client)
+
+    retry_story = ScriptedProvider.prompts[-1][1]
+    assert "First turn." in retry_story
+    assert "Second turn." not in retry_story
+
+
 def test_never_retried_action_has_no_variants(client):
     _play(client)
     last = _actions(client)[-1]