Agentforce Memory: Salesforce Doubles the Recall Rate of Facts Its AI Agents Remember

On November 3, 2026, Salesforce AI Research published on Salesforce's official site (salesforce.com/news) the detail of an update to Agentforce Memory, the feature that lets an AI agent retain what a customer or employee has already told it. The article, bylined by Itai Asseo and Egor Pakhomov, documents a measured gain in recall of memorized facts — with no change to the underlying architecture, which has been in production since September.

The Concrete Problem Salesforce Documents

Salesforce's example is deliberately simple: a customer states in March that their office is in Bangalore; they move to Pune in June; in July, the agent still answers based on Bangalore. Nothing distinguishes, from the customer's side, a reply built on an up-to-date fact from one built on a stale one — in the case Salesforce cites, a professional agent that would keep having catering delivered to a office's old address. Erik Nijkamp, director of research at Salesforce AI Research, sums up the symptom: an agent's reply never tells you on its own whether the failure came from information that was never saved, or information that was saved but poorly used.

What Changes Technically — and What Doesn't

• Agentforce Memory's architecture stays the one from its general availability launch in September — no overhaul of the storage or retrieval system.
• The change is limited to the memory policy: the rules that decide what the agent keeps from a conversation, and how it handles information that contradicts something earlier.
• The new policy retains more of what the other party says, and treats an update as a correction to an already-known fact rather than as an extra fact stored alongside the old one.
• The evaluation was run directly on the live production Agentforce system, via long, multi-session conversations generated from ConvoMem, a public benchmark Salesforce released in 2025, with facts planted and then queried several steps later.

The Figures Salesforce Measured

• 1.8 times as many facts explicitly stated by the other party correctly recalled by the agent, compared with the previous policy.
• 2.2 times as many time-varying facts (like the office move in the example) correctly recalled.
• Salesforce sums up these two combined results as end-to-end memory quality 'essentially doubled.'

What the Article Doesn't Specify

The release names no specific Salesforce edition, no region, and doesn't say whether the memory policy update applies automatically to organizations already using Agentforce Memory or requires a configuration step. It also gives no explicit year for the general availability 'in September' it mentions — get this confirmed by your Salesforce contact rather than assuming it necessarily means September 2026.

What This Actually Means for Your Deployment

• If your Agentforce agents handle information that changes over time (contact details, account status, customer preferences): the update targets precisely that case — the measured gain there (2.2x) is higher than on static facts (1.8x).
• No migration action is apparently needed on the customer side since the architecture doesn't change, but get your Salesforce contact to confirm whether the new policy is already active on your org or whether an activation step exists.
• Salesforce's evaluation method itself (testing the full path, from the customer stating a fact to the agent acting on it, rather than each component in isolation) is transferable: before blaming an AI agent for an inconsistent answer, check whether the fact was memorized before checking whether it was used.
• If you're evaluating Agentforce Memory for the first time, this change doesn't alter scope or cost — only the expected reliability on long or recurring conversations over time.

Deploying or evaluating Agentforce agents that need to retain customer information over time? Our consulting offering audits the real reliability of your setup before a large-scale rollout.