Stop retried and deleted actions from corrupting story context and memories

Three fallout bugs from keeping the retried action row alive (906ba42),
plus two long-standing cursor bugs the same investigation turned up.

Retry context leak: the row being regenerated is still attached to the
adventure, so it was replayed as established story and the model wrote a
continuation of the attempt it was meant to replace — the story visibly
blended both takes. It leaked into four places, not one: history replay,
story-card trigger matching, in-scene NPC detection, and the memory-bank
similarity query. Adds a shared context.story_actions(exclude_action_id),
threaded through build_context and retrieve_memories.

Memory holdback: a memory could summarize the just-generated turn; retry
rewrites Action.text but memory_cursor has already advanced, so the memory
was never regenerated and went on describing narration no longer in the
story. settled_story_actions() holds the newest action back one turn —
only the last action is retryable, so that makes it unreachable. The
settled list is always a prefix, so cursors stay valid and nothing is
skipped. The run_post_turn clamp deliberately still uses the full count:
clamping to settled rewinds legacy adventures a step and double-covers an
action.

Cursor bookkeeping: memory_cursor is a position into story_actions() while
Memory.source_* are Action.index values, and the two diverge as soon as
anything is deleted. Deleting a middle action slid a never-summarized
action into the covered range, skipping it forever; and pruning a memory
left the actions it covered stranded behind the cursor. Adds
note_action_removed() (called before the delete in delete_action and
undo_turn) and a rewind in prune_dangling_memories. delete_action also
now prunes at all, which it never did.

Not addressed: editing an already-summarized action still leaves its
memory stale, and the cumulative story summary can't have one fact
un-mixed from it.

117 backend tests pass, including new test_memory_settling.py (12) and
two retry-context regression tests.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UeQVy5bEjLhfgWNc27Efet

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"]

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)
 

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
 

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()