AI News Flash | Google Upgrades Gemini: AI Chatbot "Auto-Remembers" Conversations

Hello from the AI world! Today's newsletter covers the following topics:

　　・Google upgrades Gemini: AI chatbot now "auto-remembers" conversations

　　・Anthropic upgrades Claude: AI can end harmful conversations on its own for safety

　　・OpenAI expands ChatGPT's integration with popular productivity tools

GOOGLE

Google Upgrades Gemini: AI Chatbot Now "Auto-Remembers" Conversations

Google says its latest update lets the AI chatbot "Gemini" automatically remember past conversations with users, personalizing the experience without manual reminders — storing "important details and preferences" for friendlier, more practical answers.

This builds on last year's update, when users had to ask Gemini to remember preferences; now AI stores and uses it automatically. E.g., after discussing "YouTube channel ideas about Japanese culture," a later video request may prompt Gemini to suggest "Japanese food culture experiences."

As expectations for AI shift toward "closer support," this update casts AI as a "companion assistant." Google says it still prioritizes safety and privacy. Spokesperson Elijah Lawal said:
"Our goal is to build a personalized AI assistant. It's important for Gemini to understand user preferences, but we also want to give users the flexibility to turn the feature on or off at any time."

Memory is on by default; manage it under "Settings > Personal context > Past conversations with Gemini." It rolls out first on "Gemini 2.5 Pro," then "Gemini 2.5 Flash."

Google is also strengthening privacy settings. Within weeks, "Gemini Apps Activity" will be renamed "Keep Activity," letting Google use "samples" of uploaded files and images to improve its services — but only if users opt in.

Google is also adding a "Temporary Chat" mode: conversations aren't used for personalization or AI training, and data is kept at most 72 hours, letting users chat freely without leaving a "trace" — a safe space to explore, Google says.

CLAUDE AI

Anthropic Upgrades Claude: AI Can End Harmful Conversations on Its Own for Safety

Anthropic recently revealed some Claude models can now end a conversation themselves in "extremely rare" cases of harmful or prolonged user abuse — to protect the AI model, not the user.

The company isn't claiming Claude or other large language models (LLMs) are conscious or can be "hurt" by conversation, stressing that "the potential moral status of Claude and other LLMs remains highly uncertain, now and in the future."

Still, it's part of a new research program called "model welfare." Anthropic takes a precautionary approach, using "low-cost interventions" to minimize model welfare risk should the concept prove meaningful.

For now, it's limited to Claude Opus 4 and 4.1, triggering only in severe cases — e.g., requests for inappropriate content involving minors, or misuse of information enabling large-scale violence or terrorism.

Anthropic acknowledges such requests carry legal, reputational risk. In pre-release tests, Opus 4 showed a "strong tendency to refuse" them, at times with "clear signs of distress."

On the conversation-ending feature, Anthropic explains:
"Claude uses this option only as a last resort — after repeated attempts to redirect fail, when productive dialogue seems hopeless, or when the user directly asks to end the conversation."

Anthropic stressed Claude is trained never to use this on users at risk of "self-harm" or "harming others." Even after a chat ends, users can start a new one on the same account, or edit an earlier reply to branch off.

Anthropic concludes:
"This is an ongoing experiment, and we will keep refining our approach."

OPENAI

OpenAI Expands ChatGPT's Integration with Popular Productivity Tools

OpenAI rolled out an update for Pro users, letting ChatGPT connect more closely with Gmail, Google Calendar, Google Contacts, and GitHub to reference their content directly in conversations.

Plus users will also gain "connectors" for tools like Microsoft Teams and SharePoint, plus Box, Canva, Dropbox, HubSpot, Notion, and more.

As usual, Pro-tier features are expected to roll out gradually to other paid plans — Plus, Team, Enterprise, and Edu — over the coming weeks.

This marks a major leap from earlier integrations like Google Drive — the far broader connection range makes accessing, processing information faster and easier than ever.

There are limits, though: OpenAI confirms Plus/Pro connectors are unavailable in the European Economic Area (EEA), Switzerland, and the UK, and TechRadar Pro is looking into why.

This coincides with OpenAI's release of GPT-5 and GPT-5 Thinking for enterprise customers, with three modes — Auto, Fast, Thinking. Plus allows up to 3,000 GPT-5 Thinking messages weekly, then auto-switches to the lighter GPT-5 Thinking mini.

In addition, the "o4" model, briefly hidden when GPT-5 launched, has returned to the model picker. OpenAI explains:
"Paid users can enable ‘Show additional models’ in ChatGPT's web settings to show models like o3, o4-mini, 4.1, and GPT-5 Thinking mini. GPT-4.5 stays Pro-only due to GPU requirements."

Closing

We hope today's content helps with your use of AI technology and information gathering. If you found this email useful, please share it with interested colleagues or friends. We'll keep curating noteworthy topics to bring you valuable information. Look forward to our next

news

. Thank you for your continued support.