Mnema 2.2.0

DeepSeek and YandexGPT. Two more AIs by API key in Settings → AI & models. DeepSeek takes a key from platform.deepseek.com. YandexGPT takes a Yandex Cloud AI Studio key and the folder it belongs to, and brings Alice AI, YandexGPT Pro 5.1 and Pro 5, plus DeepSeek and Qwen from the same catalogue. Both get the same parts of memory as ChatGPT by default; change it in «What each AI sees».

Strata on this computer. Strata runs big models such as Qwen and DeepSeek on your own graphics card. Start a model with strata launch and Mnema finds it by itself, next to the built-in model and Ollama: free, offline, nothing leaves the computer.

Only models that answer. OpenAI lists about a hundred models, and in a chat only a third of them answer: the rest are retired (o4-mini, gpt-4, o1…), made for voice, pictures or sound, or open only through another API. When you connect ChatGPT, Gemini, DeepSeek or YandexGPT, or press «Check and find models», Mnema now asks each model one tiny question and keeps only those that reply. If a provider retires a model later, it leaves the list at its first refusal and the chat says so.

Nine languages. Arabic is gone. A saved Arabic choice switches to English, and the AI answers in English too.

README and guide in every language. The README now exists in all nine languages, linked from a table of flags. Each one opens with animations of the real interface in that language. The new Mnema guide covers every screen and setting in 42 sections, also in nine languages — the same text Mnema answers questions about itself from.

Smaller fixes.

• «Add another AI» in the consortium opens its own window and brings you back to the discussion. It used to just close.
• «Connect more» in the model menu is a list like the connected AIs above.
• The quote above a reply is a quiet card in the theme's colours, not a light grey block.