diff --git a/hindsight-api/hindsight_api/engine/consolidation/consolidator.py b/hindsight-api/hindsight_api/engine/consolidation/consolidator.py
index 276b98d8..80aafd42 100644
--- a/hindsight-api/hindsight_api/engine/consolidation/consolidator.py
+++ b/hindsight-api/hindsight_api/engine/consolidation/consolidator.py
@@ -24,9 +24,10 @@ from datetime import datetime, timezone
 from itertools import combinations
 from typing import TYPE_CHECKING, Any
 
-from pydantic import BaseModel
+from pydantic import BaseModel, field_validator
 
 from ...config import get_config
+from ..llm_wrapper import sanitize_llm_output
 from ..memory_engine import fq_table
 from ..retain import embedding_utils
 from .prompts import build_batch_consolidation_prompt
@@ -45,12 +46,22 @@ class _CreateAction(BaseModel):
     text: str
     source_fact_ids: list[str]  # memory UUIDs from the NEW FACTS list
 
+    @field_validator("text", mode="before")
+    @classmethod
+    def sanitize_text(cls, v: str) -> str:
+        return sanitize_llm_output(v) or ""
+
 
 class _UpdateAction(BaseModel):
     text: str
     observation_id: str  # UUID of the existing observation to update
     source_fact_ids: list[str]  # memory UUIDs from the NEW FACTS list
 
+    @field_validator("text", mode="before")
+    @classmethod
+    def sanitize_text(cls, v: str) -> str:
+        return sanitize_llm_output(v) or ""
+
 
 class _DeleteAction(BaseModel):
     observation_id: str  # UUID of the observation to remove
diff --git a/hindsight-api/hindsight_api/engine/consolidation/prompts.py b/hindsight-api/hindsight_api/engine/consolidation/prompts.py
index 67db5c48..17987a4b 100644
--- a/hindsight-api/hindsight_api/engine/consolidation/prompts.py
+++ b/hindsight-api/hindsight_api/engine/consolidation/prompts.py
@@ -29,7 +29,7 @@ Compare the facts against existing observations:
 - Same topic as an existing observation → UPDATE it (observation_id + source_fact_ids)
 - New topic with durable knowledge → CREATE a new observation (source_fact_ids)
 - Cross-reference facts within the batch: a later fact may resolve a vague reference in an earlier one
-- Purely ephemeral facts → omit them (no create/update needed)"""
+- Purely ephemeral facts → omit them unless the MISSION above explicitly targets such data (e.g. timestamped events, session state, screen content)"""
 
 # Output format — JSON braces escaped as {{ }} so .format() leaves them literal
 _BATCH_OUTPUT_FORMAT = """
diff --git a/hindsight-api/hindsight_api/engine/entity_resolver.py b/hindsight-api/hindsight_api/engine/entity_resolver.py
index f166e8a8..5bb1421f 100644
--- a/hindsight-api/hindsight_api/engine/entity_resolver.py
+++ b/hindsight-api/hindsight_api/engine/entity_resolver.py
@@ -459,10 +459,12 @@ class EntityResolver:
             entity_dates = [g.event_date for _, g in sorted_groups]
 
             # INSERT ... ON CONFLICT DO NOTHING — no row lock on already-existing entities.
+            # mention_count starts at 0 here; flush_pending_stats() is the sole source of
+            # truth for mention counting (one stat per original mention in the batch).
             inserted_rows = await conn.fetch(
                 f"""
                 INSERT INTO {fq_table("entities")} (bank_id, canonical_name, first_seen, last_seen, mention_count)
-                SELECT $1, name, COALESCE(event_date, now()), COALESCE(event_date, now()), 1
+                SELECT $1, name, COALESCE(event_date, now()), COALESCE(event_date, now()), 0
                 FROM unnest($2::text[], $3::timestamptz[]) AS t(name, event_date)
                 ON CONFLICT (bank_id, LOWER(canonical_name))
                 DO NOTHING
@@ -489,13 +491,15 @@ class EntityResolver:
                 for row in existing_rows:
                     id_by_name[row["name_lower"]] = row["id"]
 
-            # Assign entity IDs back and queue for post-txn stats flush.
+            # Assign entity IDs back and queue one stat per original mention so that
+            # flush_pending_stats() increments mention_count by the true mention count,
+            # not just 1 per unique name.
             for name_lower, g in sorted_groups:
                 entity_id = id_by_name.get(name_lower)
                 if entity_id:
                     for original_idx in g.indices:
                         entity_ids[original_idx] = entity_id
-                    pending.append(_EntityStat(entity_id=entity_id, event_date=g.event_date))
+                        pending.append(_EntityStat(entity_id=entity_id, event_date=g.event_date))
 
         # Accumulate into the resolver's pending list; the orchestrator flushes
         # these with await entity_resolver.flush_pending_stats() after the txn.
diff --git a/hindsight-api/hindsight_api/engine/llm_wrapper.py b/hindsight-api/hindsight_api/engine/llm_wrapper.py
index 8f1cab4d..8bfae4a8 100644
--- a/hindsight-api/hindsight_api/engine/llm_wrapper.py
+++ b/hindsight-api/hindsight_api/engine/llm_wrapper.py
@@ -48,6 +48,28 @@ _llm_max_concurrent = int(os.getenv(ENV_LLM_MAX_CONCURRENT, str(DEFAULT_LLM_MAX_
 _global_llm_semaphore = asyncio.Semaphore(_llm_max_concurrent)
 
 
+def sanitize_llm_output(text: str | None) -> str | None:
+    """
+    Sanitize text by removing characters that break downstream systems.
+
+    Removes:
+    - ASCII control characters (0x00-0x08, 0x0B-0x0C, 0x0E-0x1F, 0x7F): break
+      json.loads and PostgreSQL UTF-8 encoding; tab (0x09), newline (0x0A), and
+      carriage return (0x0D) are preserved as they are valid in text and JSON.
+    - Unicode surrogates (U+D800-U+DFFF): Invalid in UTF-8, break LLM APIs
+
+    Surrogate characters are used in UTF-16 encoding but cannot be encoded
+    in UTF-8. They can appear in Python strings from improperly decoded data
+    (e.g., from JavaScript or broken files). Control characters commonly appear
+    in LLM output embedded inside JSON string values.
+    """
+    if text is None:
+        return None
+    if not text:
+        return text
+    return re.sub(r"[\x00-\x08\x0b\x0c\x0e-\x1f\x7f\ud800-\udfff]", "", text)
+
+
 class OutputTooLongError(Exception):
     """
     Bridge exception raised when LLM output exceeds token limits.
diff --git a/hindsight-api/hindsight_api/engine/retain/fact_extraction.py b/hindsight-api/hindsight_api/engine/retain/fact_extraction.py
index c6819933..f7c897b5 100644
--- a/hindsight-api/hindsight_api/engine/retain/fact_extraction.py
+++ b/hindsight-api/hindsight_api/engine/retain/fact_extraction.py
@@ -15,7 +15,7 @@ from typing import Literal, cast
 from pydantic import BaseModel, ConfigDict, Field, create_model, field_validator
 
 from ...config import get_config
-from ..llm_wrapper import LLMConfig, OutputTooLongError
+from ..llm_wrapper import LLMConfig, OutputTooLongError, sanitize_llm_output
 from ..response_models import TokenUsage
 from .entity_labels import (
     EntityLabelsConfig,
@@ -66,25 +66,7 @@ def _infer_temporal_date(fact_text: str, event_date: datetime | None) -> str | N
 
 
 def _sanitize_text(text: str | None) -> str | None:
-    """
-    Sanitize text by removing characters that break downstream systems.
-
-    Removes:
-    - Null bytes (\\x00): Invalid in PostgreSQL UTF-8 encoding
-    - Unicode surrogates (U+D800-U+DFFF): Invalid in UTF-8, break LLM APIs
-
-    Surrogate characters are used in UTF-16 encoding but cannot be encoded
-    in UTF-8. They can appear in Python strings from improperly decoded data
-    (e.g., from JavaScript or broken files). Null bytes commonly appear in
-    OCR output, PDF extraction, or copy-paste from binary sources.
-    """
-    if text is None:
-        return None
-    if not text:
-        return text
-    # Remove null bytes and surrogate characters
-    text = text.replace("\x00", "")
-    return re.sub(r"[\ud800-\udfff]", "", text)
+    return sanitize_llm_output(text)
 
 
 class Entity(BaseModel):