Mem0 Python SDK v2.0.7 ships Gemini on Vertex AI and 25+ bug fixes
Mem0 Python SDK `v2.0.7` adds Gemini via Vertex AI as an LLM provider and native `embed_batch` support for `OllamaEmbedding`, alongside more than two dozen bug fixes across core async behavior, LLM integrations, vector stores, and embeddings.
Score breakdown
The release resolves multiple silent data-integrity bugs in async core operations — including partial deletions and dropped return values — that could corrupt memory state in production agentic applications relying on Mem0.
- 01Gemini via Vertex AI is added as a new LLM provider.
- 02Native `embed_batch` method added to `OllamaEmbedding` for batched embedding requests.
- 03`api_error_handler` was silently dropping return values from async methods — now fixed.
Mem0 Python SDK `v2.0.7` ships two feature additions alongside an extensive bug-fix pass. On the features side, Gemini via Vertex AI is now a supported LLM provider, and `OllamaEmbedding` gains a native `embed_batch` method for batched embedding requests. AWS Bedrock embeddings also gain proper `aws_session_token` support, and the Cohere and ZeroEntropy reranker fallback paths now respect `config.top_k`.
The release also fixes a `KeyError` crash when message parsers encounter messages without a `content` key, and preserves custom metadata fields during memory updates.
The core async fixes are notable in scope: `api_error_handler` was silently dropping return values from async methods, `AsyncMemory.reset()` was not resetting the entity store, and `async delete_all` was aborting on the first error and leaving memory in a partially deleted state. The release also fixes a `KeyError` crash when message parsers encounter messages without a `content` key, and preserves custom metadata fields during memory updates.
LLM provider fixes span a wide range of integrations: Anthropic's `tool_choice` format and tool response parsing are corrected; Ollama's `json` format no longer mutates the caller's messages list in-place; `None` config values are now omitted from Gemini's `GenerateContentConfig` to prevent validation errors; reasoning-model params are honored in both `OpenAIStructuredLLM` and `AzureOpenAIStructuredLLM`; `max_completion_tokens` is sent for the GPT-5 family across all providers; and `**kwargs` forwarding is added for Together, LangChain, and Sarvam providers. Vector store fixes resolve crash conditions in FAISS filtered search, Weaviate and MongoDB `reset()` calls, Pinecone hybrid search with `None` filters, Redis empty-filter crashes, and missing-ID handling in Milvus, Weaviate, Supabase, and ChromaDB `get()` calls.
Key facts
- 01Gemini via Vertex AI is added as a new LLM provider.
- 02Native `embed_batch` method added to `OllamaEmbedding` for batched embedding requests.
- 03`api_error_handler` was silently dropping return values from async methods — now fixed.
- 04`AsyncMemory.reset()` was not resetting the entity store — now fixed.
- 05`async delete_all` was aborting on first error and leaving partial deletions — now fixed.
- 06`max_completion_tokens` is now sent for the GPT-5 family across all providers.
- 07Vector store crash fixes cover FAISS, Weaviate, MongoDB, Pinecone, Redis, Milvus, Supabase, and ChromaDB.
Topics
Summary and scoring are generated automatically from the original article. We always link back to the publisher and never republish images or paywalled content. Last processed Jun 18, 2026 · 10:40 UTC. How this works →