[thalesfsp]

Summary

Today, when self.llm.generate_response raises during fact extraction in _add_to_vector_store (and its async counterpart), mem0 catches the exception, logs LLM extraction failed: {e}, and returns an empty list. The caller has no way to distinguish "LLM unavailable" from "LLM extracted no facts" — both look like [].

This makes provider-fallback wrappers impossible to build cleanly: anything wanting to retry on 429 / 5xx / connection error / timeout can't tell whether to fall through.

Change

Re-raise after logging, in both sync (_add_to_vector_store) and async (_add_to_vector_store_async) paths. The error is still logged at ERROR level for visibility.

Why this is safe

• Existing callers that don't try/except around .add() will surface previously-silent failures. This is the goal — callers should be told the LLM is down, not given a silent empty result.
• Embedding failures already propagate (no analogous swallow), so this aligns .add()'s contract: external dependency failures bubble; internal parsing/extraction-empty cases continue to return [].
• The parsing path immediately below (the second try/except) still swallows json.JSONDecodeError — that's a different class of failure (LLM responded with non-JSON) and isn't changed.

Use case driving this

Building a fallback chain in a wrapper service: primary LLM (xAI grok) → fallback (Anthropic) → fallback (local). The wrapper needs to retry .add() on the next provider when the current one errors. With the current behavior the wrapper sees an empty success and stops.

Test plan

• Existing unit tests in tests/memory/ still pass (no test currently asserts the silent-empty behavior).
• Manual: with an invalid openai_base_url, .add() previously returned []; after this change, it raises openai.APIConnectionError (or similar).

Companion PR

The same fork also has #5176 (fact-shape normalization) which is unrelated and can land independently.

🤖 Generated with Claude Code

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[blazingretinol]

closing this pr in favour of #5878

thank you for your contribution!