Salesforce Doubles AI Agent Memory with New Rules

Artificial intelligence systems often struggle to keep track of details when users update information months later. For example, if a customer mentions moving to a new office in June, an AI agent might still use the old address from March. This happens because replies based on outdated facts look identical to those based on current ones. Salesforce AI Research addressed this challenge by building a testing framework to measure how well agents remember facts across long conversations.

New Rules Fix Memory Flaws

The company used this testing setup to update its Agentforce Memory system. Instead of rebuilding the underlying technology, researchers changed the memory policy. The updated rules dictate what an agent stores and how it handles updates. Now, when a user shares new information, the system treats it as a correction rather than filing it as a separate detail. Testing showed the updated policy recalls nearly twice as many stated facts and more than double the facts that change over time.

Real-World Testing for Enterprise AI

Testing these capabilities in a live environment helps developers understand whether an agent failed to save information or simply failed to use it. Enterprise customers, such as a travel platform using the technology for internal support, benefit when agents maintain accurate context across multiple sessions. As businesses adopt these tools, testing end-to-end conversation memory is becoming an important standard for software reliability.