Mistral AI can now remember personal details about you and use them to provide better prompts. There are also new MCP connectors that businesses can deploy to connect users to third-party technical services.
Paris-based AI Biz on Tuesday offers beta versions of Le Chat’s memory, the company’s response to ChatGpt and French cat pun memory.
With data retention already available from rivals such as Openai and Anthropic, our AI chatbots can be more useful by storing details about past interactions to guide future responses.
This type of personalization brings the same potential privacy concerns as search and advertising. Careless exposure of user prompts containing personal information is already causing problems for a variety of AI services and affected users.
Perhaps because Mistral operates in Europe with substantial data regulations, we publish our privacy policy and a detailed explanation of how the data is used above and the options for customers to control it.
“When you enter sensitive data into the input, such as health details, this data can be stored as memory to provide more relevant and personalized answers,” the company’s documentation explains, saying that memory is an opt-in service.
As an example, Biz tells Le Chat that he has a peanut allergy and suggests, “Le Chat may remember it to rule out peanuts from his recipe suggestions.”
Note that there is a big difference between “may remember” and “remember.” The company’s posts on the subject suggest that Le Chat has an 86% chance of getting accurately stored information.
That’s not exactly certain. Therefore, those with a peanut allergy may want to think a lot about what Le Chat trusts to order takeout (or you can place “no peanuts” at the restaurant prompt).
This is now a slightly more plausible scenario thanks to the release of the “20+ Secure MCP-Driven Connector” that allows Le Chat customers to connect to business-oriented tools.
MCP stands for Model Context Protocol. This is how developers can enable AI models to interact with third-party services. Therefore, empowered AI models, often referred to as “agents,” do not function very well.
If the word “secure” really applies here, it’s epic. However, there is a lack of previous MCP implementations, such as Anthropic’s SQLite MCP server and the MCP service of the AI code editor Cursor. Security company Pynt recently found that one in ten MCP plugins is fully exploitable, and having three such plugins increases the risk of over 50% of exploitability.
Nevertheless, Mistral declares that “admin users can confidently control the connectors available to anyone in their organization, and Ombharf authentication allows users to access only the data they are permitted to.”
Available connectors include Asana, Atlassian, Box, Brevo, Cloudflare, Databricks (coming soon), Github, Linear, Monday.com, Concepts, PayPal, Pinecone, Plaid, Prisma, Postgres, Salesforce Soon), Sentry, Snowflake (Soon), Square, Stripe, and Zapier. ®
