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:
     """