fix(oss): normalize malformed LLM fact output before embedding (#4224)

Co-authored-by: kartik-mem0 <kartik.labhshetwar@mem0.ai>

mem0/memory/main.py

@@ -29,6 +29,7 @@ from mem0.memory.utils import (
     ensure_json_instruction,
     extract_json,
     get_fact_retrieval_messages,
+    normalize_facts,
     parse_messages,
     parse_vision_messages,
     process_telemetry_filters,
@@ -456,6 +457,7 @@ class Memory(MemoryBase):
                     # Try extracting JSON from response using built-in function
                     extracted_json = extract_json(response)
                     new_retrieved_facts = json.loads(extracted_json)["facts"]
+                new_retrieved_facts = normalize_facts(new_retrieved_facts)
         except Exception as e:
             logger.error(f"Error in new_retrieved_facts: {e}")
             new_retrieved_facts = []
@@ -1484,6 +1486,7 @@ class AsyncMemory(MemoryBase):
                     # Try extracting JSON from response using built-in function
                     extracted_json = extract_json(response)
                     new_retrieved_facts = json.loads(extracted_json)["facts"]
+                new_retrieved_facts = normalize_facts(new_retrieved_facts)
         except Exception as e:
             logger.error(f"Error in new_retrieved_facts: {e}")
             new_retrieved_facts = []

mem0/memory/utils.py

@@ -1,12 +1,15 @@
 import hashlib
+import logging
 import re
 
 from mem0.configs.prompts import (
+    AGENT_MEMORY_EXTRACTION_PROMPT,
     FACT_RETRIEVAL_PROMPT,
     USER_MEMORY_EXTRACTION_PROMPT,
-    AGENT_MEMORY_EXTRACTION_PROMPT,
 )
 
+logger = logging.getLogger(__name__)
+
 
 def get_fact_retrieval_messages(message, is_agent_memory=False):
     """Get fact retrieval messages based on the memory type.
@@ -77,6 +80,30 @@ def format_entities(entities):
 
     return "\n".join(formatted_lines)
 
+def normalize_facts(raw_facts):
+    """Normalize LLM-extracted facts to a list of strings.
+
+    Smaller LLMs (e.g. llama3.1:8b) sometimes return facts as objects
+    like {"fact": "..."} or {"text": "..."} instead of plain strings.
+    This mirrors the TypeScript FactRetrievalSchema validation.
+    """
+    if not raw_facts:
+        return []
+    normalized = []
+    for item in raw_facts:
+        if isinstance(item, str):
+            fact = item
+        elif isinstance(item, dict):
+            fact = item.get("fact") or item.get("text")
+            if fact is None:
+                logger.warning("Unexpected fact shape from LLM, skipping: %s", item)
+                continue
+        else:
+            fact = str(item)
+        if fact:
+            normalized.append(fact)
+    return normalized
+
 
 def remove_code_blocks(content: str) -> str:
     """

tests/test_memory.py

@@ -1,9 +1,11 @@
+import json
 from unittest.mock import MagicMock, patch
 
 import pytest
 
 from mem0 import Memory
 from mem0.configs.base import MemoryConfig
+from mem0.memory.utils import normalize_facts
 
 
 class MockVectorMemory:
@@ -244,4 +246,77 @@ def test_get_all_handles_flat_list_from_postgres(mock_sqlite, mock_llm_factory,
 
     assert len(result) == 2
     assert result[0]["memory"] == "Memory 1"
-    assert result[1]["memory"] == "Memory 2" 
+    assert result[1]["memory"] == "Memory 2"
+
+
+@patch('mem0.utils.factory.EmbedderFactory.create')
+@patch('mem0.utils.factory.VectorStoreFactory.create')
+@patch('mem0.utils.factory.LlmFactory.create')
+@patch('mem0.memory.storage.SQLiteManager')
+def test_add_infer_with_malformed_llm_facts(mock_sqlite, mock_llm_factory, mock_vector_factory, mock_embedder_factory):
+    """
+    Repro for: 'list' object has no attribute 'replace' on infer=true.
+
+    When an LLM (especially smaller models like llama3.1:8b) returns facts as
+    objects ({"fact": "..."} or {"text": "..."}) instead of plain strings,
+    the embedding model's .replace() call crashes with AttributeError.
+    """
+    mock_embedder = MagicMock()
+    mock_embedder.embed.side_effect = lambda text, action: (_ for _ in ()).throw(
+        AttributeError("'dict' object has no attribute 'replace'")
+    ) if not isinstance(text, str) else [0.1, 0.2, 0.3]
+    mock_embedder_factory.return_value = mock_embedder
+
+    mock_vector_store = MagicMock()
+    mock_vector_store.search.return_value = []
+    mock_vector_factory.return_value = mock_vector_store
+
+    # LLM returns malformed facts: dicts instead of strings
+    malformed_response = json.dumps({
+        "facts": [
+            {"fact": "User likes Python"},
+            {"text": "User is a developer"},
+        ]
+    })
+    mock_llm = MagicMock()
+    mock_llm.generate_response.return_value = malformed_response
+    mock_llm_factory.return_value = mock_llm
+
+    mock_sqlite.return_value = MagicMock()
+
+    from mem0.memory.main import Memory as MemoryClass
+    config = MemoryConfig()
+    memory = MemoryClass(config)
+
+    # This should NOT raise AttributeError
+    memory._add_to_vector_store(
+        messages=[{"role": "user", "content": "I like Python and I'm a developer"}],
+        metadata={"user_id": "test_user"},
+        filters={"user_id": "test_user"},
+        infer=True,
+    )
+
+
+def test_normalize_facts_plain_strings():
+    assert normalize_facts(["fact one", "fact two"]) == ["fact one", "fact two"]
+
+
+def test_normalize_facts_dict_with_fact_key():
+    assert normalize_facts([{"fact": "User likes Python"}]) == ["User likes Python"]
+
+
+def test_normalize_facts_dict_with_text_key():
+    assert normalize_facts([{"text": "User is a developer"}]) == ["User is a developer"]
+
+
+def test_normalize_facts_mixed():
+    raw = [
+        "plain string",
+        {"fact": "from fact key"},
+        {"text": "from text key"},
+    ]
+    assert normalize_facts(raw) == ["plain string", "from fact key", "from text key"]
+
+
+def test_normalize_facts_filters_empty_strings():
+    assert normalize_facts(["", "valid", ""]) == ["valid"]