Memory
What an agent remembers between conversations, how that memory is scoped, and how it differs by runtime.
On this page
Your agent can build long-term memory from conversations, so it can recall relevant context in future interactions. How it remembers depends on the agent's runtime.
| Runtime | How it remembers |
|---|---|
| Anthropic and OpenAI Responses | Extracts facts from conversations into shared and personal memory, as described on this page |
| Claude Agent SDK | Keeps its own memory as files in the agent's workspace |
| Classic | Recalls related messages from the agent's earlier conversations |
Anthropic and OpenAI Responses agents#
This section also applies to legacy Upstage agents.
How it works#
As the agent converses with users, it extracts factual information, such as which team owns a service, where a document lives, or what tool is used for a specific workflow. These facts are stored and recalled in future conversations to provide more contextual responses.
The agent extracts "resource → location" style facts only: concrete, referenceable information. It does not store tasks, opinions, or personal data.
Shared and personal memory#
Each fact is stored as shared or personal memory, depending on where the conversation happens.
Shared memory is available to the agent in all of its conversations, so any user benefits from it. For example, if someone mentions in a public channel that "the design system lives in the ui-kit repo," the agent remembers this for everyone.
Personal memory is tied to one user. For example, if a user tells the agent in a direct message "I work on the billing team," that user's future private conversations include this context.
| Where the conversation happens | Memory |
|---|---|
| Slack direct messages and private channels, for messages posted there | Personal |
| Slack @mentions (including in private channels), reactions, and slash commands | Shared |
| Microsoft Teams personal chats and group chats | Personal |
| Microsoft Teams channels | Shared |
| Discord | Shared |
| The dashboard playground | Shared |
When memory is written#
The agent adds to memory from conversations in Slack, Microsoft Teams, Discord, and the dashboard playground. It does not add to memory from HubSpot or Zendesk tickets, from triggers, or from runs started only because a watched user was mentioned. Requests through the REST API and the Web SDK are treated as private but carry no end-user identity, so they don't add to memory either.
Claude Agent SDK agents#
Agents on the Claude Agent SDK runtime keep memory as files in the agent's own workspace, which they read and update as they work. The Enable long-term memory toggle doesn't govern this memory.
Classic agents#
Classic agents recall related messages from the agent's earlier conversations. This recall is scoped to the agent as a whole, with no split between personal and shared memory. Turning off Enable long-term memory turns this recall off.
Settings#
Memory is controlled by the Enable long-term memory toggle in your agent's settings ("Remember user preferences and frequently referenced resources across conversations"). It is on by default.
For Anthropic and OpenAI Responses agents, turning it off stops the agent from extracting and recalling memories, but does not delete existing ones.