Double recall: Inside the Agentforce Memory update

Salesforce AI Research developed an evaluation framework to assess how much an agent remembers by generating multi-session conversations with embedded facts and testing the agent’s responses in real time. The framework, called ConvoMem, was used to improve Agentforce Memory, which became generally available in September. The goal was to address the challenge of agents retaining outdated information, such as a customer’s old office location, which could lead to incorrect actions like misdelivered catering orders.

The updated memory policy in Agentforce Memory retains more customer statements and treats updates as corrections rather than additional facts. This change was implemented without altering the underlying architecture, demonstrating that policy adjustments alone could significantly enhance performance. Testing on the live Agentforce system showed the new policy recalled 1.8 times as many stated facts and 2.2 times as many updated facts compared to the previous policy.

Erik Nijkamp, director of research at Salesforce AI Research, emphasized that testing the entire conversation path is essential to distinguish between failures in saving information and failures in using it. This approach allows for targeted improvements by adjusting policies rather than rebuilding systems. Silvio Savarese, EVP and chief scientist at Salesforce, highlighted the importance of honest measurement in production environments, where real-world conditions like long conversations and live data are critical.

A Salesforce customer, a travel platform, implemented an employee-facing agent using Agentforce to answer questions, summarize cases, draft emails, and update CRM records. With the improved Agentforce Memory, the agent now carries relevant context into every conversation, providing personalized and accurate responses without requiring users to repeat information.