How did you know that?!?

I asked my AI assistant why a stock was down, and its answer implied it knew details about my portfolio I hadn't explicitly shared in that session. That unexpected 'memory' sent me down a rabbit hole....

I wanted help with my personal investments, so I started researching different investment funds, stocks, bonds, you name it.  I didn’t actually load specific details into the LLM but I did ask it to compare investment opportunities and told it, in generic terms, some of my goals. 

Things were going well until I was in a brand new chat session and asked why a particular stock was down for the day.  To my surprise, in the model’s answer, it implied I owned the stock (based on my researching it earlier and asking questions about the income it would generate) and gave me some overall advice for my portfolio.  Where did it get that????  So, I asked.

LLM Chat Session

After several probing prompts, the model conceded that it can, sometimes, infer that something is important enough to remember.  Yet again it reminded me of how important my privacy is and offered to not remember things unless I specifically asked.

Memory is important.  Without it the model’s responses wouldn’t be very helpful, but this clearly demonstrates that it isn’t clear or obvious what the model remembers, why, or how long it remembers, raising significant issues around data privacy, transparency, and user control..  ChatGPT announced on April 10 that it will now be able to “reference all of your past chats to provide more personalized responses, drawing on your preferences and interests”.  While enhanced memory can be helpful, this lack of explicit user control is precisely where the risks lie, especially for businesses handling sensitive information. This challenge is central to what we're solving at Cognita.

The Cognita Platform™ protects you, and your company, giving you direct control over conversational memory. By treating the underlying LLMs as powerful but stateless engines, Cognita ensures your context is managed securely and transparently according to your rules  Your information is stored in encrypted form and never used for any purposes outside of your company. Cognita™ has many other features that protect your data and ideas while ensuring that you get the most out of the LLM. 

This is just the beginning. Follow Cognita as we share more insights into building trustworthy AI systems and unveil the features that put you in control of your data."

Interested in how secure, controlled AI memory can benefit your enterprise? Visit www.cognitaai.net to learn more about our vision, sign up for our newsletter, and access our blog.