diff --git a/python/packages/core/agent_framework/_harness/_memory.py b/python/packages/core/agent_framework/_harness/_memory.py
index c3f02f883b..731af85092 100644
--- a/python/packages/core/agent_framework/_harness/_memory.py
+++ b/python/packages/core/agent_framework/_harness/_memory.py
@@ -108,19 +108,44 @@ def _normalize_memory_text(memory: str) -> str:
     return normalized
 
 
+_LEADING_BACKSLASHES_BEFORE_HASH: re.Pattern[str] = re.compile(r"^(\\+)#")
+
+
 def _escape_markdown_line(line: str) -> str:
-    """Escape a markdown line so it cannot be re-interpreted as a heading on a later read."""
-    if line.lstrip().startswith(("#", "\\#")):
-        return f"\\{line}"
+    """Escape a markdown line so it cannot be re-interpreted as a heading on a later read.
+
+    A line starting with ``#`` gets one backslash. A line that already starts
+    with backslashes before ``#`` gets one more than it had, so a reader can
+    always recover the exact original run: prepending a single backslash to an
+    escaped hash produced a two backslash prefix the old reader never stripped,
+    corrupting the round trip.
+    """
+    leading_whitespace_length = len(line) - len(line.lstrip())
+    body = line[leading_whitespace_length:]
+    match = _LEADING_BACKSLASHES_BEFORE_HASH.match(body)
+    if match:
+        run_length = len(match.group(1))
+        escaped_run = "\\" * (2 * run_length + 1)
+        return line[:leading_whitespace_length] + escaped_run + body[run_length:]
+    if body.startswith("#"):
+        return "\\" + line
     return line
 
 
 def _unescape_markdown_line(line: str) -> str:
-    """Reverse :func:`_escape_markdown_line` when reading persisted memory content."""
+    """Reverse :func:`_escape_markdown_line` when reading persisted memory content.
+
+    An odd backslash run before ``#`` is one this module wrote (``n`` originals
+    become ``n + 1``), so the reader keeps ``(run - 1) // 2`` originals. An even
+    run was not produced here and is left untouched.
+    """
     leading_whitespace_length = len(line) - len(line.lstrip())
     body = line[leading_whitespace_length:]
-    if body.startswith("\\#"):
-        return f"{line[:leading_whitespace_length]}{body[1:]}"
+    match = _LEADING_BACKSLASHES_BEFORE_HASH.match(body)
+    if match and len(match.group(1)) % 2 == 1:
+        original_length = (len(match.group(1)) - 1) // 2
+        original_run = "\\" * original_length
+        return line[:leading_whitespace_length] + original_run + body[len(match.group(1)) :]
     return line
 
 
diff --git a/python/packages/core/tests/core/test_harness_memory.py b/python/packages/core/tests/core/test_harness_memory.py
index 464255d0b4..df33cb2d5f 100644
--- a/python/packages/core/tests/core/test_harness_memory.py
+++ b/python/packages/core/tests/core/test_harness_memory.py
@@ -1046,6 +1046,24 @@ def test_memory_file_store_encodes_non_ascii_owner_ids(tmp_path) -> None:
     assert roots[0] != roots[1]
 
 
+def test_memory_topic_record_round_trips_backslash_hash_lines() -> None:
+    """Lines starting with backslash hash must survive a round trip unchanged.
+
+    Escaping such a line used to add a second backslash the reader never
+    stripped, so every reload corrupted the memory a little more.
+    """
+    record = MemoryTopicRecord(
+        topic="notes",
+        summary="\\# summary line",
+        memories=["\\# literal fact", "# heading shaped memory", "\\\\# two backslashes"],
+        updated_at="2026-10-26T00:00:00+00:00",
+    )
+
+    parsed = MemoryTopicRecord.from_markdown(record.to_markdown(), fallback_topic="notes")
+
+    assert parsed.memories == ["\\# literal fact", "# heading shaped memory", "\\\\# two backslashes"]
+
+
 def test_memory_file_store_uses_literal_folders_for_safe_identifiers(tmp_path) -> None:
     """Safe identifiers are readable on disk rather than opaque encoded segments."""
     store = MemoryFileStore(tmp_path, owner_prefix="user_", owner_state_key="owner_id")