Mistral AI can now remember personal details about you and use them to offer better prompts. It also has new MCP connectors that businesses can deploy to connect their users to third-party tech services.
The Paris-based AI biz on Tuesday is now offering a beta version of Memories for Le Chat, the company's answer to ChatGPT and a French-language feline pun.
The data retention, already available from rivals like OpenAI and Anthropic, allows the company's AI chatbot to be more helpful by storing stated preferences and details about past interactions to guide future responses.
Personalization of this sort poses the same potential privacy concern as it does for search and advertising. The inadvertent exposure of user prompts containing personal info has already posed problems for various AI services and affected users.
Perhaps because Mistral operates out of Europe where there's substantive data regulation, the company has published a detailed explanation in its Privacy Policy and documentation about how it might use said data and the options customers have to control it.
"If you include sensitive data in your Input, such as health details, this data may be stored as a Memory to provide you with more relevant and personalized answers," the company's documentation explains, noting that Memories is an opt-in service.
As an example, the biz suggests that if you tell Le Chat you have a peanut allergy, "Le Chat may remember it in order to exclude peanuts from recipe suggestions."
We note that there's a significant difference between "may remember" and "will remember." The company's post on the subject suggests Le Chat's chance of accurately retrieving saved information is 86 percent.
That's not exactly a sure thing. So those with peanut allergies might want to think twice about trusting Le Chat to order takeout (or they could just put "no peanuts" in the restaurant prompt).
That's a slightly more plausible scenario now, thanks to Mistral's release of "20+ secure, MCP-powered connectors" that allow Le Chat customers to connect to business-oriented tools.
MCP stands for Model Context Protocol. It's a way for developers to allow AI models to interact with third-party services. AI models empowered thus are often referred to as "agents," though they haven't worked all that well.
If the word "secure" truly applies here, that would be grand. But MCP implementations to date, such as Anthropic's SQLite MCP server or AI code editor Cursor's MCP service, have fallen short. Security firm Pynt recently found that one in 10 MCP plugins is fully exploitable and having three such plugins raised the risk of exploitability beyond 50 percent.
Mistral nonetheless proclaims that "Admin users can confidently control which connectors are available to whom in their organization, with on-behalf authentication, ensuring users only access data they're permitted to."
Available connectors include: Asana, Atlassian, Box, Brevo, Cloudflare, Databricks (coming soon), GitHub, Linear, Monday.com, Notion, PayPal, Pinecone, Plaid, Prisma, Postgres, Salesforce soon), Sentry, Snowflake (soon), Square, Stripe, and Zapier. ®