After parsing LLM JSON response, normalize facts before passing them
to the embedding model. Fixes 'list'/'dict' object has no attribute
'replace' when smaller LLMs return malformed fact objects.
Port of TypeScript FactRetrievalSchema to Python.
Normalizes facts that smaller LLMs return as {"fact": "..."} or
{"text": "..."} objects back into plain strings before embedding.
When smaller LLMs return facts as objects ({"fact": "..."}) instead of
plain strings, embedding_model.embed() crashes with
'dict' object has no attribute 'replace'. This test confirms the bug
by mocking the LLM to return dict-shaped facts and asserting the
failure at mem0/memory/main.py:473.
anishamahuli