Salesforce Reports Agentforce Memory Gains

Salesforce AI Research says an updated policy for Agentforce Memory recalls 1.8 times as many facts customers state outright and 2.2 times as many facts that change over time as the policy it replaced. The team measured the update with an evaluation framework run on the live Agentforce system. Agentforce Memory became generally available in September, though the article does not state the year.

The framework generates long, multi-session conversations with specific facts inserted along the way, runs them through Agentforce, then asks questions that depend on those facts. It tests the full path from when a customer shares information to when an agent acts on it. The conversations come from ConvoMem, a public benchmark Salesforce released in 2025.

The updated memory policy changes which customer details an agent keeps and how it handles new information. It retains more of what customers say and treats an update as a correction to an existing fact, rather than storing it beside the earlier fact. Salesforce says the underlying Agentforce Memory architecture has not changed. The evaluation found the end-to-end conversation memory was essentially doubled in quality.

Erik Nijkamp, director of research at Salesforce AI Research, said an agent’s answer alone cannot show whether it failed to save information or failed to use information it saved. Testing the whole path helps identify which failure occurred, he said, allowing the team to change a policy instead of rebuilding the system. Salesforce says production testing matters because memory operates across months of conversations, live data and users waiting for answers.

A travel platform customer used Agentforce to build an employee-facing agent that answers team questions, summarizes cases, drafts emails and updates CRM records. Salesforce says improved Agentforce Memory carries relevant context into each conversation, giving sales and service representatives personalized answers without starting from scratch. The company argues businesses will hold agent memory to standards like other parts of their systems, and says testing full conversations is a way to set that standard.

Salesforce AI Research says an updated policy for Agentforce Memory recalls 1.8 times as many facts customers state outright and 2.2 times as many facts that change over time as the policy it replaced. The team measured the update with an evaluation framework run on the live Agentforce system. Agentforce Memory became generally available in September, though the article does not state the year.

The framework generates long, multi-session conversations with specific facts inserted along the way, runs them through Agentforce, then asks questions that depend on those facts. It tests the full path from when a customer shares information to when an agent acts on it. The conversations come from ConvoMem, a public benchmark Salesforce released in 2025.

The updated memory policy changes which customer details an agent keeps and how it handles new information. It retains more of what customers say and treats an update as a correction to an existing fact, rather than storing it beside the earlier fact. Salesforce says the underlying Agentforce Memory architecture has not changed. The evaluation found the end-to-end conversation memory was essentially doubled in quality.

Erik Nijkamp, director of research at Salesforce AI Research, said an agent’s answer alone cannot show whether it failed to save information or failed to use information it saved. Testing the whole path helps identify which failure occurred, he said, allowing the team to change a policy instead of rebuilding the system. Salesforce says production testing matters because memory operates across months of conversations, live data and users waiting for answers.

A travel platform customer used Agentforce to build an employee-facing agent that answers team questions, summarizes cases, drafts emails and updates CRM records. Salesforce says improved Agentforce Memory carries relevant context into each conversation, giving sales and service representatives personalized answers without starting from scratch. The company argues businesses will hold agent memory to standards like other parts of their systems, and says testing full conversations is a way to set that standard.