Agentforce Memory update doubles recall of changing facts by rewriting only the memory policy, not the architecture
Synopsis
Salesforce AI Research updated the memory policy behind Agentforce Memory so an agent treats a customer's later statement as a correction to a known fact rather than a second fact filed beside it. The evaluation ran on the live Agentforce system, using long multi-session conversations from the public ConvoMem benchmark with specific facts planted along the way, then asked the agent questions that depend on those facts. 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, with nothing underneath the policy changed.
Interpretation
An agent now holds on to information a customer revised months earlier, and the gain is quantified: after Salesforce AI Research updated the memory policy behind Agentforce Memory, the agent recalls 1.8 times as many facts customers state outright as the policy it replaced. Even today's most advanced agents can miss details a customer updates months later, because a reply built on an older fact reads exactly like a reply built on a current one, leaving the agent unable to tell them apart. The comparison comes from an evaluation run on the live Agentforce system, with conversations generated by ConvoMem, a public benchmark Salesforce released in 2025, facts planted along the way, and questions then asked about those facts.
For facts that change over time, the updated policy recalls 2.2 times as many as the policy it replaced, essentially doubling the memory quality of the end-to-end conversation. The earlier policy treated a customer's new statement as an additional fact filed next to the first, rather than as a correction to what the agent already knew. Both multipliers come from the same evaluation on the live Agentforce system, covering the whole path from the moment a customer mentions something to the moment an agent acts on it.
The improvement sits in the memory policy alone, the rules that decide what an agent keeps about a customer and what it does when that information changes; Agentforce Memory still runs on the architecture it launched with. The updated policy holds on to more of what a customer says and treats an update as a correction to something the agent already knows. Salesforce states that nothing underneath the policy changed, so the recall change can be attributed to the policy itself.
Perspective
The demonstrated scope is an evaluation on the live Agentforce system using multi-session conversations generated by ConvoMem, plus one usage setting in which a travel platform built an employee-facing agent with Agentforce; the material proposes no next experiment or deployment plan.
The recall multipliers are comparisons against the prior policy, and the material does not state absolute recall levels; the evaluation ran on a live system, and the material stresses that memory only behaves like memory under months of conversation, live data and a human waiting on the answer.