fix(memory): re-raise LLM extraction failures instead of returning [] (salvage of #5178) (#5878)

mem0/memory/main.py

@@ -22,6 +22,7 @@ from mem0.configs.prompts import (
     PROCEDURAL_MEMORY_SYSTEM_PROMPT,
     generate_additive_extraction_prompt,
 )
+from mem0.exceptions import LLMError
 from mem0.exceptions import ValidationError as Mem0ValidationError
 from mem0.memory.base import MemoryBase
 from mem0.memory.setup import mem0_dir, setup_config
@@ -912,8 +913,12 @@ class Memory(MemoryBase):
                 response_format={"type": "json_object"},
             )
         except Exception as e:
+            # Re-raise so callers can implement provider fallback / retry.
+            # The original silent ``return []`` made upstream callers unable to
+            # distinguish "LLM unavailable" (429/5xx/timeout) from "LLM
+            # extracted no facts" -- both surfaced as an empty list.
             logger.error(f"LLM extraction failed: {e}")
-            return []
+            raise LLMError(f"LLM extraction failed: {e}") from e
 
         # Parse response
         try:
@@ -2536,8 +2541,10 @@ class AsyncMemory(MemoryBase):
                 response_format={"type": "json_object"},
             )
         except Exception as e:
+            # Re-raise so callers can implement provider fallback / retry
+            # (see sync counterpart for rationale).
             logger.error(f"LLM extraction failed (async): {e}")
-            return []
+            raise LLMError(f"LLM extraction failed: {e}") from e
 
         # Parse response
         try:

tests/memory/test_main.py

@@ -6,6 +6,7 @@ from unittest.mock import MagicMock, Mock
 
 import pytest
 
+from mem0.exceptions import LLMError
 from mem0.memory.main import AsyncMemory, Memory
 
 
@@ -79,6 +80,28 @@ class TestAddToVectorStoreErrors:
         assert mock_memory.llm.generate_response.call_count == 1
         assert result == []  # Should return empty list when no memories processed
 
+    def test_llm_extraction_exception_is_reraised(self, mocker, mock_memory):
+        """A provider error during fact extraction must propagate, not be swallowed.
+
+        Regression guard for the silent ``return []`` that made it impossible for
+        callers to tell "LLM unavailable" (429/5xx/timeout) from "no facts found".
+        Without the fix this raises AssertionError because the call returns [].
+        """
+
+        class _ProviderError(Exception):
+            pass
+
+        mock_memory.llm.generate_response.side_effect = _ProviderError("429 rate limit")
+        mocker.patch("mem0.memory.main.capture_event")
+
+        with pytest.raises(LLMError) as exc_info:
+            mock_memory._add_to_vector_store(
+                messages=[{"role": "user", "content": "test"}], metadata={}, filters={}, infer=True
+            )
+        # The documented LLMError contract is honoured, and the original
+        # provider exception is preserved as the cause for debugging.
+        assert isinstance(exc_info.value.__cause__, _ProviderError)
+
 
 class TestPromptOverridesCustomInstructions:
     @pytest.fixture
@@ -245,6 +268,29 @@ class TestAsyncAddToVectorStoreErrors:
         assert result == []
         assert mock_async_memory.llm.generate_response.call_count == 1
 
+    @pytest.mark.asyncio
+    async def test_async_llm_extraction_exception_is_reraised(self, mock_async_memory, mocker):
+        """Async counterpart of the sync re-raise guard.
+
+        A provider error during fact extraction must propagate as ``LLMError``
+        (with the original exception preserved as the cause), not be swallowed
+        into ``return []``. Without the fix a future revert of the async
+        ``raise`` back to ``return []`` would pass the suite silently.
+        """
+        mocker.patch("mem0.utils.factory.EmbedderFactory.create", return_value=MagicMock())
+
+        class _ProviderError(Exception):
+            pass
+
+        mock_async_memory.llm.generate_response.side_effect = _ProviderError("429 rate limit")
+        mocker.patch("mem0.memory.main.capture_event")
+
+        with pytest.raises(LLMError) as exc_info:
+            await mock_async_memory._add_to_vector_store(
+                messages=[{"role": "user", "content": "test"}], metadata={}, effective_filters={}, infer=True
+            )
+        assert isinstance(exc_info.value.__cause__, _ProviderError)
+
 
 def _build_memory_instance(mocker, memory_cls):
     _setup_mocks(mocker)