Salesforce says policy update doubled Agentforce Memory recall

The problem: agents that act on outdated facts

Salesforce illustrates the issue with a simple scenario. In March, a customer tells an agent their office is in Bangalore. In June, they say it has moved to Pune. In July, they ask a question that depends on location, and the agent answers as if the office were still in Bangalore.

According to Salesforce, keeping details current when a customer updates them months later is one of the harder problems in AI agent memory. Even today's most advanced agents can miss such changes, because a reply built on an older fact reads exactly like one built on a current fact. The company says the consequences range from a customer having to enter information twice to an employee's catering order being delivered to the old office.

How Salesforce measured memory and what changed

Salesforce AI Research built an evaluation framework meant to answer how much an agent actually remembers. It generates long, multi-session conversations with specific facts planted along the way, runs them through the live Agentforce system, and then asks the agent questions that depend on those facts. The conversations come from ConvoMem, a public benchmark Salesforce released in 2025. Testing the entire path, from the moment a customer mentions something to the moment an agent acts on it, is what makes the result trustworthy, according to the company.

The framework was used to improve Agentforce Memory, the capability that keeps agents current on what customers tell them. Salesforce says it became generally available in September. The gain came from an updated memory policy, the rules that decide what an agent keeps and what it does when information changes. The new policy holds on to more of what a customer says. It also treats an update as a correction to something the agent already knows, rather than an additional fact filed next to the first.

Salesforce stresses that nothing underneath the policy changed, and that Agentforce Memory still runs on the architecture it launched with. On the evaluation, run on the live system, the updated policy recalls 1.8 times as many facts customers state outright as the policy it replaced, and 2.2 times as many facts that change over time.

Erik Nijkamp, director of research at Salesforce AI Research, said: "An agent's answer can't tell you whether it failed to save something or failed to use what it saved." He added that once you can see which one is failing, "you can improve it by changing a policy instead of rebuilding the system."

Why production testing matters for customers

Salesforce argues that evaluating in production is important because memory only behaves like memory under real conditions: months of conversation, live data, and a human waiting on the answer. Silvio Savarese, EVP and chief scientist at Salesforce, said that in a product organization "the problem is already running, with customers on it, and the discipline is to measure it honestly before you change anything."

The company points to one customer, a travel platform, that used Agentforce to build an employee-facing agent. The agent answers team questions, summarizes cases, drafts emails and updates CRM records. With the improved memory, Salesforce says, the agent carries the relevant context into every conversation, so sales and service reps get personalized, relevant answers without starting from scratch each time.

Salesforce expects agent memory to become something enterprise businesses hold to a standard, as they do with any other part of their systems. In its view, testing what an agent remembers across a full conversation is how that standard gets set. The reported figures come from Salesforce's own evaluation, and the source does not describe independent verification.