MemCut tests if AI agents remove a secret when a user tells them to
After 10 more messages were removed, a reader got the secret from the remaining data at a rate of 39 to 76 percent.
Claimed, not confirmed
Researchers made MemCut, a test for AI agents that keep a long-term memory of user conversations. The test has 456 secrets that the authors made. It checks if an agent removes a secret when a user tells it to. The team tested 5 agent memory systems and 4 language models. No different group checked the result. If the message that tells the secret is removed, the usual retrieval loses much more than a reader that gets the remaining data.
Sources
Posted
