fix(memory): skip text-less turns when formatting the memory prompt (#7985)

* fix(memory): skip text-less turns when formatting the memory prompt

A content-less message rendered as the literal "None", and a block list
without text blocks rendered as its Python repr, including any base64
data URL. Normalize through the same rules as the sibling content-to-text
helpers and skip turns that carry no text.

* test(memory): cover whitespace-only assistant turns

---------

Co-authored-by: Willem Jiang <willem.jiang@gmail.com>

backend/packages/harness/deerflow/agents/memory/backends/deermem/deermem/core/prompt.py

@@ -759,6 +759,33 @@ def format_memory_for_injection(
     return result
 
 
+def _content_to_text(content: Any) -> str:
+    """Normalize LangChain message content for the memory prompt.
+
+    Mirrors ``deerflow/utils/messages.py:message_content_to_text``, the mem0
+    ``message_filtering`` copy, and ``message_processing.extract_message_text``:
+    a content-less message is empty rather than the truthy literal ``"None"``,
+    and a block list without text blocks contributes no text. ``str(content)``
+    on such a list would embed its Python repr in the extraction prompt,
+    including a ``data:`` URL's base64 payload for an image-only turn.
+    """
+    if content is None:
+        return ""
+    if isinstance(content, str):
+        return content
+    if isinstance(content, list):
+        text_parts = []
+        for p in content:
+            if isinstance(p, str):
+                text_parts.append(p)
+            elif isinstance(p, dict):
+                text_val = p.get("text")
+                if isinstance(text_val, str):
+                    text_parts.append(text_val)
+        return " ".join(text_parts)
+    return str(content)
+
+
 def format_conversation_for_update(messages: list[Any]) -> str:
     """Format conversation messages for memory update prompt.
 
@@ -771,27 +798,17 @@ def format_conversation_for_update(messages: list[Any]) -> str:
     lines = []
     for msg in messages:
         role = getattr(msg, "type", "unknown")
-        content = getattr(msg, "content", str(msg))
-
-        # Handle content that might be a list (multimodal)
-        if isinstance(content, list):
-            text_parts = []
-            for p in content:
-                if isinstance(p, str):
-                    text_parts.append(p)
-                elif isinstance(p, dict):
-                    text_val = p.get("text")
-                    if isinstance(text_val, str):
-                        text_parts.append(text_val)
-            content = " ".join(text_parts) if text_parts else str(content)
+        content = _content_to_text(getattr(msg, "content", str(msg)))
 
         # Strip the upload-context tag from human messages to avoid persisting
         # ephemeral file path info into long-term memory.  Skip the turn entirely
-        # when nothing remains after stripping (upload-only message).
+        # when nothing remains after stripping (upload-only message). Turns with
+        # no text at all (content-less AI message, image-only blocks) are skipped
+        # the same way instead of rendering a bare "Assistant: " line.
         if role == "human":
-            content = re.sub(r"<current_uploads>[\s\S]*?</current_uploads>\n*", "", str(content)).strip()
-            if not content:
-                continue
+            content = re.sub(r"<current_uploads>[\s\S]*?</current_uploads>\n*", "", content).strip()
+        if not content.strip():
+            continue
 
         # Truncate very long messages: keep the head (topic / opening) and the
         # tail (conclusion / "remember X" instruction), dropping the middle.

backend/tests/test_memory_updater.py

@@ -1093,6 +1093,47 @@ class TestFormatConversationForUpdate:
         assert "Tom & Jerry" in result
         assert "a < b" in result
 
+    def test_content_less_message_is_skipped_instead_of_rendering_none(self):
+        """A content-less message must not become the literal "None" turn.
+
+        ``str(None)`` is truthy, so the previous fallback fed ``Assistant: None``
+        to the extraction prompt as if the assistant had said it. Sibling helpers
+        (``utils/messages.py:message_content_to_text``, the mem0 copy, and
+        ``message_processing.extract_message_text``) all return "" here.
+        """
+        msg = MagicMock()
+        msg.type = "ai"
+        msg.content = None
+
+        assert format_conversation_for_update([msg]) == ""
+
+    def test_whitespace_only_assistant_turn_is_skipped(self):
+        msg = MagicMock()
+        msg.type = "ai"
+        msg.content = "   "
+
+        assert format_conversation_for_update([msg]) == ""
+
+    def test_image_only_blocks_do_not_leak_the_block_repr(self):
+        """An image-only turn contributes no text, not a base64 Python repr."""
+        payload = "data:image/png;base64," + "A" * 2000
+        msg = MagicMock()
+        msg.type = "human"
+        msg.content = [{"type": "image_url", "image_url": {"url": payload}}]
+
+        assert format_conversation_for_update([msg]) == ""
+
+    def test_text_blocks_render_when_an_image_block_is_also_present(self):
+        msg = MagicMock()
+        msg.type = "human"
+        msg.content = [
+            {"type": "text", "text": "what is in this screenshot?"},
+            {"type": "image_url", "image_url": {"url": "data:image/png;base64,AAAA"}},
+        ]
+
+        result = format_conversation_for_update([msg])
+        assert result == "User: what is in this screenshot?"
+
 
 # ---------------------------------------------------------------------------
 # update_memory - structured LLM response handling