Description

When calling POST /api/v1/memories/ with infer=true, smaller LLMs (e.g. Ollama's llama3.1:8b) intermittently return facts as objects ({"fact": "..."} or {"text": "..."}) instead of plain strings. These non-string values are passed directly to embedding_model.embed(), which calls .replace("\n", " ") on them, causing AttributeError: 'list' object has no attribute 'replace'.

Fixes #4100

Root cause

The fact extraction path in mem0/memory/main.py parses the LLM's JSON response and passes each fact directly to the embedding model without validating its type:

new_retrieved_facts = json.loads(response)["facts"]
# ...
for new_mem in new_retrieved_facts:
    messages_embeddings = self.embedding_model.embed(new_mem, "add")  # assumes string

The prompt asks the LLM to return {"facts": ["string1", "string2"]}, but smaller models don't reliably follow this format. Instead they return structures like:

{"facts": [{"fact": "User likes Python"}, {"text": "User is a developer"}]}

This is a known issue — the TypeScript SDK already fixed it in db15d5c6 by adding a FactRetrievalSchema that normalizes these malformed shapes before embedding. See mem0-ts/src/oss/src/prompts/index.ts. The Python SDK was missing equivalent validation.

Fix

This PR ports that validation to Python by adding normalize_facts() in mem0/memory/utils.py, which handles:

• Plain strings (passthrough)
• {"fact": "..."} objects (extracts the fact value)
• {"text": "..."} objects (extracts the text value)
• Other types (converts via str())
• Empty strings (filtered out)

It is called in both the sync and async _add_to_vector_store paths immediately after JSON parsing, before facts reach the embedder.

Related issues: #3439, #3238

Type of change

Please delete options that are not relevant.

• Bug fix (non-breaking change which fixes an issue)

How Has This Been Tested?

Please describe the tests that you ran to verify your changes. Provide instructions so we can reproduce. Please also list any relevant details for your test configuration

Please delete options that are not relevant.

• Unit Test
  hatch run pytest tests/test_memory.py — 17/17 pass.

Tests added:

• Reproduction test (test_add_infer_with_malformed_llm_facts): mocks LLM returning dict-shaped facts with infer=True, confirms no AttributeError.
• normalize_facts unit tests: plain strings, {"fact": ...}, {"text": ...}, mixed lists, empty string filtering

Checklist:

• My code follows the style guidelines of this project
• I have performed a self-review of my own code
• I have commented my code, particularly in hard-to-understand areas
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes
• I have checked my code and corrected any misspellings

Maintainer Checklist

• closes OpenMemory infer=true fails with 'list' object has no attribute 'replace' (Qdrant path) #4100
• Made sure Checks passed