From 3ae1120403c1b9bc286a01a7f1c3a1ad35a223e9 Mon Sep 17 00:00:00 2001
From: anishamahuli <anisha.mahuli@gmail.com>
Date: Thu, 5 Mar 2026 12:12:53 -0500
Subject: [PATCH] Add normalize_facts() utility for malformed LLM fact
 extraction output

Port of TypeScript FactRetrievalSchema to Python.
Normalizes facts that smaller LLMs return as {"fact": "..."} or
{"text": "..."} objects back into plain strings before embedding.
---
 mem0/memory/utils.py | 21 ++++++++++++++++++++-
 1 file changed, 20 insertions(+), 1 deletion(-)

diff --git a/mem0/memory/utils.py b/mem0/memory/utils.py
index 8c11705c8..6cf4a477b 100644
--- a/mem0/memory/utils.py
+++ b/mem0/memory/utils.py
@@ -2,9 +2,9 @@ import hashlib
 import re
 
 from mem0.configs.prompts import (
+    AGENT_MEMORY_EXTRACTION_PROMPT,
     FACT_RETRIEVAL_PROMPT,
     USER_MEMORY_EXTRACTION_PROMPT,
-    AGENT_MEMORY_EXTRACTION_PROMPT,
 )
 
 
@@ -52,6 +52,25 @@ 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.
+    """
+    normalized = []
+    for item in raw_facts:
+        if isinstance(item, str):
+            fact = item
+        elif isinstance(item, dict):
+            fact = item.get("fact") or item.get("text") or str(item)
+        else:
+            fact = str(item)
+        if fact:
+            normalized.append(fact)
+    return normalized
+
 
 def remove_code_blocks(content: str) -> str:
     """