Agentforce Memory policy update recalls 1.8x more stated facts and 2.2x more changed facts

Salesforce AI Research says an updated memory 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, compared with the policy it replaced. According to Salesforce, the gains came without changing the architecture Agentforce Memory launched with.

What changed

Agentforce Memory is the capability that keeps agents current on what customers tell them. Salesforce says it became generally available in September. The improvement comes from the memory policy: the rules that decide what an agent keeps about a customer and what it does when that information changes.

According to Salesforce, the updated policy does two things differently:

• It holds on to more of what a customer says.
• It treats an update as a correction to something the agent already knows, rather than as an additional fact filed next to the first.

Salesforce's example: a customer says in March that their office is in Bangalore, then says in June that it moved to Pune. In July, an agent with stale memory still answers as though the office were in Bangalore. Salesforce notes that a reply built on an older fact reads exactly like one built on a current fact, which makes these failures hard to spot.

How it was measured

Salesforce AI Research built an evaluation framework that generates long, multi-session conversations with specific facts planted along the way. The conversations are run through the live Agentforce system, and the agent is then asked questions that depend on those facts. The conversations come from ConvoMem, a public benchmark Salesforce released in 2025.

The company says this tests the entire path, from the moment a customer mentions something to the moment an agent acts on it. On this evaluation, Salesforce reports:

• 1.8x the recall of facts customers state outright, versus the previous policy.
• 2.2x the recall of facts that change over time, versus the previous policy.

Salesforce describes the result as the memory of the end-to-end conversation being essentially doubled in quality. Absolute recall rates were not published in the article.

Statements from Salesforce

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. 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."

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."

Customer use

Salesforce says a travel platform customer recently used Agentforce to build an employee-facing agent that answers team questions, summarizes cases, drafts emails and updates CRM records. With the improved memory, the company says, the agent carries relevant context into each conversation so sales and service reps do not start from scratch.

What this means

The notable claim is where the improvement came from: a policy change, not a new architecture. Salesforce's framing is that separating "failed to save" from "failed to use what was saved" is what makes a policy-level fix possible, and the end-to-end test on the live system is how it tells the two apart.

For teams building agents with long-running customer context, the example of a changed office address shows why updates are handled as corrections rather than additional facts. The 1.8x and 2.2x figures are Salesforce's own results on its own evaluation, built on its ConvoMem benchmark, so they describe improvement over its previous policy and not a comparison with other systems.