Default Gemini model (gemini-2.0-flash) was retired June 1, Gemini provider fails unless a model is explicitly set

[benminor]

Component

Python SDK

Description

Summary

Both mem0 SDKs (Python and TS) fall back to gemini-2.0-flash when no Gemini model is specified. Google retired the Gemini 2.0 Flash family from the API on June 1, 2026 (https://ai.google.dev/gemini-api/docs/deprecations), so the Gemini provider fails out of the box unless the user explicitly overrides the model.

Fallback locations:

• mem0/llms/gemini.py:37 --> if not self.config.model: self.config.model = "gemini-2.0-flash"
• mem0-ts/src/oss/src/llms/google.ts:13 --> same default in the TS SDK
• mem0/configs/llms/gemini.py:35 --> docstring recommends the retired ID

Steps to Reproduce

from mem0 import Memory

config = {
    "llm": {
        "provider": "gemini",
        "config": {"api_key": "<GOOGLE_API_KEY>"},  # no model specified
    }
}
m = Memory.from_config(config)
m.add("I prefer window seats", user_id="u1")

Expected Behavior

Configuring the Gemini provider with just an API key works --> the default model is one Google currently serves.

Actual Behavior

The request is rejected by the Gemini API with a model-not-found / deprecated-model error, because the hardcoded fallback gemini-2.0-flash was retired June 1, 2026. (Tests pin the same ID but mock the API calls, so CI stays green through the retirement.)

Related, not yet failing: the default OpenAI model rides the Oct 23, 2026 shutdown (https://platform.openai.com/docs/deprecations) in several load-bearing spots — server/main.py:116 (DEFAULT_LLM_MODEL → gpt-4.1-nano-2025-04-14), mem0/configs/llms/base.py:34, mem0/llms/base.py:59 (o1/o3-mini), the dashboard defaults (server/dashboard/src/app/setup/page.tsx:446, .../configuration/page.tsx:152), and ~15 examples/ files. Flagging now since one sweep is cheaper than incident-driven fixes in October.

Happy to open a PR for the Gemini fallback if maintainers pick a target model (Google's docs point at the 2.5/3.x line).

Found by running arol, a deprecation scanner I built after a vendor retirement broke our own production - filing because the Gemini half is live breakage.

Environment

• mem0 version: current main (cloned 2026-07-29)
• Python/Node version: N/A - the failing ID is hardcoded in source
• OS: N/A

[Solaris-star]

+1 on this — the Gemini fallback is live breakage, and I agree with the reporter's instinct to wait for a maintainer-chosen target model rather than guess.

One small note for whoever picks this up: the reporter flagged that the OpenAI default (gpt-4.1-nano-2025-04-14) rides the Oct 23, 2026 shutdown across server/main.py, mem0/configs/llms/base.py, mem0/llms/base.py, the dashboard defaults, and ~15 examples. Doing the Gemini fix and the OpenAI sweep in one pass would be cheaper than incident-driven fixes in October. Happy to help with the PR once maintainers confirm the target models.

[O96a]

The default should be gemini-2.5-flash — it's the current recommended general-purpose model, same price class as 2.0 Flash was, and Google's been stable on it since launch. The TS SDK default at google.ts:13 needs the same swap.

While you're in there, the OpenAI defaults flagged in the report (gpt-4.1-nano-2025-04-14 riding the Oct shutdown) are worth fixing in the same pass since they touch the same config layer. I'd bump those to gpt-4.1-nano (the latest stable nano) — OpenAI keeps the short name pointing at the current recommended version.

Happy to review a PR if you open one.

[ratishjain12]

Would like to raise a PR for this if no one is working on it currently.

[benminor]

Would like to raise a PR for this if no one is working on it currently.

No worries have already made the necessary changes, opening PR shortly. Thanks all!