Agent memory correction doubled what an enterprise agent recalled. On 4 November, Salesforce AI Research said an updated memory policy for Agentforce Memory recalled 1.8 times as many of the facts customers state outright, and 2.2 times as many of the facts that change over time. The architecture underneath did not change. The policy did. An update is now treated as a correction to what the agent already knows, rather than as an additional fact filed next to the first.

Salesforce opens with a failure any team can test. A customer says in March that an office is in Bangalore. In June, the customer says the office moved to Pune. In July, the agent still answers from Bangalore. A reply built on an old fact reads exactly like a reply built on a current one, Salesforce notes, and the cost can be concrete: catering ordered to an office the company left.

Agent memory correction beat a rebuild

The result matters because nothing below the policy was rebuilt. Agentforce Memory, which became generally available in September, still runs on the architecture it launched with, according to Salesforce. The policy holds on to more of what a customer says, and it changes what an update means. Appending preserves a contradiction and leaves retrieval to choose between two versions with no stated winner. Correction names a winner.

That is the same fault line a company record faces at larger scale. A supplier term, an approval limit or a delivery address does not sit in one conversation. It governs answers to many people and, increasingly, actions by other agents. If the old version stays live beside the new one, every later retrieval becomes a coin toss the reader cannot see.

The measurement behind the claim is end to end. Salesforce said its framework generates long, multi session conversations with facts planted along the way, runs them through the live Agentforce system, and then asks questions that depend on those facts. The conversations come from ConvoMem, the public benchmark Salesforce released in 2025. The underlying ConvoMem preprint holds 75,336 question and answer pairs across user facts, assistant recall, abstention, preferences, temporal changes and implicit connections.

The test separates saving from using

"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, in the post. "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."

ConvoMem also supplies the caution. In the preprint, simple full context approaches scored 70 to 82 percent on the most challenging multi message cases, while RAG based memory systems such as Mem0 scored 30 to 45 percent on histories under 150 interactions. A memory layer, in other words, can score below the raw conversation it summarises. Salesforce's November figures are a vendor evaluation, run by Salesforce on its own live system with its own framework. The post names no independent replication, and its one customer example, a travel platform, carries no figure that separates the memory gain from the rest of the deployment. Read the doubling as a measured policy effect inside one system, not as a market wide benchmark.

The correction question is now a product surface elsewhere. On 3 November, Snowflake said Memory in CoCo and CoWork remembers past context and decisions, and that users can view, update or delete stored memories at any time. The same release put CoWork in more than 5,800 accounts and CoCo in more than 9,100. Sycamore, which made its Build platform generally available on 3 November, asks employees to provide organizational context and review implementation plans before software is built from that context. Different products, one shared admission: somebody has to be able to change what the system believes.

The Correction Test for a company record

Correction has a risk of its own. An update can be wrong, or narrower than it sounds. An office move for one team is not an office move for the company. Automatic correction that erases the old version would trade one silent error for another. For a company record, correction therefore needs four properties, which together form the Correction Test.

First, an update replaces in effect, not in history. The new fact governs the next answer, and the superseded version stays inspectable with its source and date. Second, the correction names its scope: the team, region or period it applies to, so a local change does not rewrite a global fact. Third, the correction propagates. Derived summaries, skills and queued actions built on the old fact are flagged or rebuilt, not left running on a version the record has already retired. Fourth, the test is run end to end, the Salesforce lesson: plant a fact, change it, then ask a fresh question that depends on the change, and save the failed cases.

Each step maps to how Jilu is described on this site. The How Jilu Works pillar sets out a record that keeps source, date and superseded version visible, with a review gate before a draft joins the shared record. Yesterday's post on the fresh session test for persistent agent memory asked whether a correction survives a new session. The earlier Memory Tax post asked whether a memory system beats the raw source at all. The Correction Test joins those questions: replace in effect, keep in history, bound the scope, and propagate.

Companies do not lose the truth because nobody stored it. They lose it at the moment a fact changes and the system files the change beside the original, with no winner named. Salesforce doubled recall by naming the winner. A company brain earns trust the same way, with one addition a customer chat does not need: the losing version stays on the record, dated and sourced, so a person can see what changed and why the answer changed with it.