fix: drop new_memories that echo Hollow's own reply (recall pollution)

On recall turns Hollow retells a memory in first person; extraction
sometimes captured that poetic line back as a new treasure entry.
Deterministic filter: >60% token overlap with the reply = Hollow's
words, not the visitor's.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

app.py

@@ -313,7 +313,7 @@ def chat(user_msg: str, state: dict, history: list):
         turn_failed = True
 
     treasure_before = len(state["treasure"])
-    state = apply_update(state, raw_json)
+    state = apply_update(state, raw_json, reply=reply)
     captured = len(state["treasure"]) > treasure_before
 
     if do_recall and recall_memory and not turn_failed:

memory.py

@@ -17,7 +17,21 @@ def get_tier(affinity: int) -> str:
     return "Almost"
 
 
-def apply_update(state: dict, raw_json: str) -> dict:
+def _is_hollows_words(memory: str, reply: str) -> bool:
+    """On recall turns Hollow retells a memory in first person; extraction
+    sometimes captures THAT poetic line as a new memory. Drop any candidate
+    whose words substantially overlap Hollow's reply (>60% shared tokens)."""
+    if not reply:
+        return False
+    mem_words = set(re.findall(r"\w+", memory.lower()))
+    if not mem_words:
+        return False
+    reply_words = set(re.findall(r"\w+", reply.lower()))
+    overlap = len(mem_words & reply_words) / len(mem_words)
+    return overlap > 0.6
+
+
+def apply_update(state: dict, raw_json: str, reply: str = "") -> dict:
     try:
         start = raw_json.find("{")
         end = raw_json.rfind("}") + 1
@@ -31,7 +45,7 @@ def apply_update(state: dict, raw_json: str) -> dict:
 
     state["affinity"] = max(0, min(100, state["affinity"] + update.affinity_delta))
     for mem in update.new_memories:
-        if mem not in state["treasure"]:
+        if mem not in state["treasure"] and not _is_hollows_words(mem, reply):
             state["treasure"].append(mem)
     state["tone"] = max(-100, min(100, state.get("tone", 0) + update.tone_delta))
     if update.cruel_quote and update.tone_delta < 0:

tests/test_memory.py

@@ -324,3 +324,27 @@ class TestSingleGateBranchesByTone:
     def test_gate_at_70_branches_only_by_tone(self):
         assert decide_ending(_end_state(affinity=70, tone=20)) == "good"
         assert decide_ending(_end_state(affinity=70, tone=19)) == "loop"
+
+
+class TestHollowWordsFiltered:
+    def _s(self):
+        return {"affinity": 50, "treasure": [], "claimed": [], "tone": 0, "wounds": []}
+
+    def test_drops_memory_that_echoes_hollows_reply(self):
+        reply = "i remember a voice, soft like the wind through the trees."
+        s = apply_update(self._s(),
+                         '{"affinity_delta": 3, "new_memories": ["a voice soft like the wind through the trees"]}',
+                         reply=reply)
+        assert s["treasure"] == []
+
+    def test_keeps_genuine_visitor_memory(self):
+        reply = "that sounds lovely. tell me more about her."
+        s = apply_update(self._s(),
+                         '{"affinity_delta": 3, "new_memories": ["had a grandmother who kept bees"]}',
+                         reply=reply)
+        assert s["treasure"] == ["had a grandmother who kept bees"]
+
+    def test_no_reply_keeps_everything(self):
+        s = apply_update(self._s(),
+                         '{"affinity_delta": 3, "new_memories": ["grew up near the sea"]}')
+        assert s["treasure"] == ["grew up near the sea"]