Note
Access to this page requires authorization. You can try signing in or changing directories.
Access to this page requires authorization. You can try changing directories.
Note
Memory is rolling out for general availability. Memory might still be in preview in some regions.
Memory lets an agent carry what it learns into later conversations, such as a person's role, preferences, and corrections. People don't need to repeat themselves, and the agent builds on what it already knows instead of starting over.
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.
What an agent remembers
An agent keeps selected facts, preferences, and lessons from the tasks it performs to help improve the quality and efficiency of future runs:
- About a person: Role, expertise, and preferences. For example, "Procurement lead for EMEA. Prefers comparisons as tables."
- Feedback: Corrections and confirmed approaches, with the reason. For example, "Score delivery terms before price, because late deliveries cost more."
- Ongoing work: Goals, constraints, progress, and next steps. For example, "For BID-105, the user requires delivery by December 15 and human approval before awarding the contract."
- References: Where to find information. For example, "Supplier ratings are in the Vendor Scorecard dashboard."
- Running records: Amounts, durations, or counts kept as dated entries.
The agent decides what to keep, guided by built-in rules and the maker's instructions. It keeps specific, useful facts, not a record of every conversation. Learn more in How agents determine what to remember.
An agent keeps memories separately for each user account and can recall them in conversations or automated runs under that account. Each person can view only their own memories.
Note
Agents that have their own account use shared memory instead of personal memory. Learn more in Memory for agents with an account.
How agents apply memory
Memories are short plain-text notes. The agent treats them as background information, not as instructions.
- Checks memory first: At the start of a conversation, the agent reads memories related to the request.
- Rechecks when a topic returns: Responses reflect what's saved.
- Favors the current conversation: If a person contradicts a memory, the agent follows the conversation and corrects or removes the outdated memory.
- Verifies before acting: If a memory points to a file, resource, or setting, the agent confirms that it still exists.
- Doesn't invent memories: If nothing relevant is saved, the agent says so.
Memory compared with conversation history and knowledge
Memory is one of several sources of information that an agent draws on.
| Source | What it provides |
|---|---|
| Memory | Context from earlier conversations, such as roles, preferences, and ongoing projects |
| Conversation history | What was said earlier in the current conversation |
| Knowledge | Reference information, such as policies and documents |
| Tools | Current data, and actions in other systems |
Memory is separate from conversation history. Within a long conversation, the agent can refer to earlier messages without saving anything, so a response that reflects earlier context doesn't mean that a memory exists.
Memory isn't a source of authority. It can't grant permissions or bypass approvals, and it can be wrong or out of date. For policies, permissions, and current status, the agent relies on knowledge and tools.