THE APEX TIMES
The company says a new evaluation framework and a revised memory policy can materially improve how its AI agents handle customer information that changes over time.
Salesforce said it has improved Agentforce Memory, the capability intended to keep AI agents aligned with what customers say across extended conversations. The update is aimed at a common failure mode in customer-facing assistants: when a person corrects or updates a detail weeks or months later, an agent may still act as though the older information is current, producing friction and potentially expensive mistakes.
In an example Salesforce uses to illustrate the problem, a customer tells an agent in March that an office is in Bangalore. In June, the customer says the office moved to Pune. Later, in July, when the customer asks a question that depends on the office location, the agent answers as if the office were still in Bangalore. Salesforce frames the underlying cause as the agent’s memory, where older facts can persist and new statements may be handled in a way that does not fully override prior context.
To quantify how well agents actually remember and how those memories get used, Salesforce AI Research built an evaluation framework designed around difficult-to-game end-to-end testing. Rather than evaluating memory in isolation, the framework generates long, multi-session conversations with specific facts intentionally placed along the way. Those conversations are then run through Salesforce’s live Agentforce system, and the agent is tested with follow-on questions that depend on the planted facts. Salesforce says this approach is meant to be trustworthy specifically because it tests the full path from a customer making a statement to the agent taking an action on it.
Salesforce said the framework draws from “ConvoMem,” a public benchmark the company released in 2025. ConvoMem is positioned as a yardstick for measuring agent memory, and the company said the new framework is meant to answer a question other methods struggle with: how much an agent really remembers when information changes over time.
The company used the evaluation framework to improve Agentforce Memory by updating the “memory policy,” the set of rules that governs what the agent keeps about a customer and what the agent does when information changes. Salesforce said the updated policy holds on to more of what a customer says and, crucially, treats an update as a correction to something the agent already knows rather than simply adding the new statement alongside earlier facts. Salesforce also emphasized that the improvement is policy-driven, not a rewrite of the underlying Agentforce Memory architecture.
On the company’s evaluation run on the live Agentforce system, Salesforce reported that the updated policy recalls 1.8 times as many facts that customers state outright, and 2.2 times as many facts that change over time, compared with the policy it replaced. Salesforce described the result as effectively doubling the quality of memory across an end-to-end conversation, given that both what is stated and what evolves are tested.
Salesforce officials said the ability to distinguish between failures to save information and failures to use what was saved is what makes the evaluation actionable. “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. He added that testing the whole path helps identify which component is failing, enabling improvements through policy changes rather than rebuilding the system. Salesforce’s chief scientist, Silvio Savarese, said the discipline matters because the problem exists in production, with real customers and real answers to wait on, rather than being measured only under lab conditions.
Agentforce Memory itself is described as something that became generally available in September, and Salesforce said the update reinforces that the capability has to behave like memory under real conditions, including months of conversation and live data. The company also offered an example of usage in a travel platform: the customer used Agentforce to build an employee-facing agent that can answer team questions, summarize cases, draft emails, and update CRM records. Salesforce said that with improved memory, the agent carries forward the context that matters into every conversation, reducing the need for reps to restart context-gathering each time.
Salesforce did not provide additional detail on the size of the benchmark, the specific metrics used beyond the reported recall multipliers, or how the policy changes affect different types of corrections. The company also did not disclose whether the improved memory behavior is available as part of a specific pricing tier or whether there are configuration options for enterprise customers. As with many AI-assistant improvements, the practical impact will likely vary with how customers phrase updates and how frequently critical facts change.
Next, Salesforce is likely to face expectations around transparency and reliability as Agentforce Memory becomes a more standard component of enterprise AI workflows. Buyers and operators will want to see continued evaluation that focuses on corrections over time, along with evidence that the improved recall translates into fewer operational errors, such as booking to the wrong location or acting on outdated customer preferences.
Apple’s December-quarter gross margin outlook, framed around rising memory costs, is landing as a potential benefit for investors, Jim Cramer said, pointing to how slowing data center build-outs could ease a “memory squeeze” in the chip supply chain.
Oracle shares have faced fresh uncertainty after a reported cancellation of a large chip-leasing arrangement involving Tencent, attributed to pressure from the Trump administration, shortly after the company raised $20 billion.
A report tying chip orders to expected delivery schedules suggests Chinese AI and cloud customers, including Alibaba, are getting a small share of RTX Pro 5500 allocations despite placing large orders.
The company says it will invest heavily to expand Google data-center capacity and support regional AI services in Malaysia, according to a report cited by Yahoo Finance.
A sharp downgrade in investor expectations sent Intel’s stock down nearly half in a single session, as the semiconductor company warned of a weaker near-term outlook and fueled a broader chip-sector selloff.
A report says Amazon has restricted Meta’s Muse from making purchases on its platform. The same coverage points to Walmart and Shopify as more open to shopping agents that can help users discover and buy products through conversational AI.
AI Markets Group estimates the market for software that approves, constrains and logs enterprise AI use will grow sharply through the decade, citing Microsoft, Databricks and Salesforce among leading vendors.
Advanced Micro Devices said it has taken on (or is underwriting) customer-facing commitments tied to AI data-center builds that could total $8.5 billion, according to a market report. The move underscores the intensifying chip competition in the spending race for artificial intelligence infrastructure.
Shares of TeraWulf fell again even after reporting indicated Google had put a long-term guarantee behind Anthropic leases at the company’s Kentucky campus, while smaller crypto-and-AI power peers sank further.
Alphabet’s Google says actor Henry Cavill will use Googlebook AI features to create a playable 16-bit game, with proceeds supporting men’s health through Movember.