Salesforce AI Research doubled Agentforce memory performance by creating a framework that tracks what an AI agent remembers.
Imagine the following scenario: In March, a customer tells an agent that their office is in Bangalore. In June, they mention the office has moved to Pune. In July, they ask a question that depends on where they are, but the agent answers as though the office were still in Bangalore. The culprit is the agent’s memory.
Keeping details current when a customer updates it months later is one of the harder problems in AI agent memory. Even today’s most advanced agents can miss them, because a reply built on an older fact reads exactly like a reply built on a current one. At best, this means a customer has to enter information twice. At worst, an employee asks the agent to arrange catering for tomorrow’s meeting and it books delivery to the old office. Either way, an agent’s “forgetfulness” translates to friction in the experience – with potential for real business consequences.
Salesforce AI Research built an evaluation framework that answers a question that is difficult to answer any other way: how much does an agent actually remember? Our framework 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. 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. The conversations come from ConvoMem, a public benchmark we released in 2025.
That framework was used to improve Agentforce Memory, the capability that keeps agents current on what customers tell them, which became generally available in September.
The improvement came from our updated memory policy, the rules that decide what an agent keeps about a customer and what it does when that information changes. The updated policy holds on to more of what a customer says, and it treats an update as a correction to something the agent already knows rather than an additional fact filed next to the first. What’s significant about this methodology is that nothing underneath the policy changed. Agentforce Memory still runs on the architecture it launched with.
On that evaluation, run on the live Agentforce system, the updated policy recalls 1.8 times as many of the facts customers state outright as the policy it replaced, and 2.2 times as many of the facts that change over time. The memory of the end-to-end conversation is essentially doubled in quality.
“An agent’s answer can’t tell you whether it failed to save something or failed to use what it saved,” said Erik Nijkamp, director of research at Salesforce AI Research. “Testing the whole path is what separates those, and once you can see which one is failing, you can improve it by changing a policy instead of rebuilding the system.”
Doing that evaluation in production matters, 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: “In a research lab, you choose a problem and then go find data for it. 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. That is what this work gives us, on a capability our customers depend on every day.”
One Salesforce customer, a travel platform, recently used Agentforce to create an employee-facing agent that answers team questions, summarizes cases, drafts emails, and updates CRM records. With its improved Agentforce Memory, the agent carries the context that matters into every conversation, so sales and service reps get personalized, relevant answers without starting from scratch each time.
Agent memory is becoming something enterprise businesses will hold to a standard, the way they hold any other part of their systems. Testing what an agent remembers across a full conversation is how that standard gets set.
