How agents determine what to remember

Note

Memory is rolling out for general availability. Memory might still be in preview in some regions.

This article explains how the agent saves memories, the criteria it uses, the information it stores, and the information it excludes.

Agents that have their own account use shared memory. Learn more in Memory for agents with an account.

Note

The harness used to build this agent uses usage-based billing. Using, building, testing, and evaluating agents might consume Copilot Credits. Learn more in Overview of usage-based billing.

How agents save memories

An agent saves memories in two ways:

  • On request: A user can say, "Remember that I manage the Main Street store." The agent still evaluates whether the information is worth retaining, so a request doesn't guarantee a save. If the request concerns something that matters only to the current conversation, the agent asks what about it is worth remembering, and saves that.
  • Automatically: The agent saves useful context during conversations and triggered runs, without a separate request to remember. For example, a revised bid prompts a purchasing agent to recommend Supplier B instead of Supplier A. The agent can remember why the recommendation changed: Supplier B now meets the delivery deadline at a lower price, and human approval is still pending.

Examples of memory decisions

The following table shows examples of how an agent decides what to remember and what to exclude:

What happens How the agent might handle it
"Remember that I manage the Main Street store." The agent saves it at the user's request.
"I work from the Seattle office. Book a room for Friday's review." The agent might save the office location automatically. It doesn't save the room booking, which is a one-time task.
"No, show that as a table." The agent saves it as feedback for future responses.
"I've moved to the APAC team." The agent updates the saved role instead of adding a duplicate.
A revised quote triggers a new review of BID-105. Supplier B now meets the delivery deadline at USD 90,000, below Supplier A's USD 100,000. The agent can update memory with the new recommendation, the reason for the change, and the pending human decision. Supplier B's earlier rejection remains as history, not the current assessment.
"Just this once, send the report to my manager." The agent doesn't save it as a preference.
"What's the travel expense policy?" The agent doesn't save it. It's a general question, not information about the user.

When saving information about a person or project, the agent is instructed to turn relative dates such as "next Friday" into specific dates. For temporary arrangements, provide a date range, such as "I'm working from the regional office October 12–16." Dates describe when the information applies; they don't set an automatic deletion date.

Memory selection criteria

Saving a memory involves judgment, so the agent might handle similar situations differently. The agent evaluates:

  • Relevance: The information helps the agent understand the person or carry out its responsibilities. "I manage onboarding for our Berlin office" provides context for future help. "What is employee onboarding?" asks for general information, not a fact to remember.
  • Durability: The information remains useful beyond the current conversation or run, which includes lasting preferences, roles, goals, and corrections, as well as project progress needed for a later review.
  • Eligibility: The agent is instructed not to save excluded information, even on request.
  • Novelty: The agent updates an existing memory instead of adding a duplicate or conflicting one, and deletes memories that prove to be incorrect.

To guide the agent to remember only facts that users state directly, add an instruction. Learn more in Guide memory with agent instructions.

Information excluded from memory

For memories kept for an individual, the agent is instructed not to save:

  • Passwords, credentials, secrets, or access tokens.
  • Anything a user asks it not to remember.
  • Personal details about other people, such as ID numbers, birth dates, account or card numbers, phone numbers, and addresses.
  • Information shared in confidence.
  • Content from other users.
  • Details that matter only to the current conversation.

These rules guide the agent's judgment. They aren't a filter that blocks content, and they complement your organization's data-protection controls rather than replace them. If the agent saves information that it shouldn't, users can ask the agent to forget it or delete it in the memory portal.