This system learns logical rules from old chats to answer new questions better.

Chatbots that talk with people over long periods struggle to answer questions about past chats. Most systems simply store old details as passive facts, which leads to weak reasoning and missed connections across huge conversation histories. To fix this, researchers built RuleMem. Instead of just saving raw details, RuleMem pulls reusable logic rules out of earlier chats to actively guide how it finds evidence and reasons. It creates these rules as natural-language if-then statements, called Horn clauses, and tests them using a check called Rule Perplexity Consistency. These rules help the system connect distant clues and generate answers. When tested on the LoCoMo benchmark against 14 other systems, RuleMem earned the highest accuracy, beating the baseline average by 27.47 points.

Why it matters

For people using long-term chatbots like personal assistants, this means answers that connect past and present more accurately. It reduces errors where the system forgets or misremembers important details from earlier conversations.

Questions answered