Short answer: shared team memory means a Slack AI agent can use what one teammate told it when it answers someone else. OpenAI's help article says the ChatGPT app in Slack, @ChatGPT, does not record or access memory at launch, and your personal ChatGPT memories do not automatically carry over to it (OpenAI Help Center). Unlike the ChatGPT app in Slack, a Runbear agent on the Anthropic or OpenAI Responses runtime keeps resource facts from Slack @mentions, such as where a document lives, as shared memory for the whole team. Facts from direct messages stay personal.
When you choose a shared AI agent for a team, check whether it keeps shared memory and what goes into it. Runbear runs shared AI agents for teams in Slack and Microsoft Teams, and on its Anthropic and OpenAI Responses runtimes, a resource fact one person gives the agent in a Slack @mention can shape any teammate's later answers. Each Runbear agent runs on a runtime, the engine chosen when you create it, and memory works differently on each. Runbear's recommended Standard option usually gives you a Claude Agent SDK agent, which keeps memory as files in its own workspace. This post compares how each option remembers. For a side-by-side list of AI agents you can add to Slack, see Runbear's best AI agents for Slack comparison.
Personal AI memory vs shared team memory in Slack
You probably first met AI memory in a personal assistant. ChatGPT Memory, for example, remembers preferences and details from your own chats. You review it under Settings, Personalization, Memory, and the controls vary by plan and workspace settings (OpenAI Help Center). That memory belongs to one person.
The ChatGPT app in Slack is a different product with different rules. Your organization's @ChatGPT runs through a service account configured for the assistant, not through your personal ChatGPT account. It does not automatically inherit your access to files, apps or private channels, and it works from the conversation context available to it and the tools your organization connects. Your saved memories, custom instructions and ChatGPT conversation history do not carry over to @ChatGPT automatically, so you include the context it needs in each request. At launch, @ChatGPT does not record or access memory (Using @ChatGPT in Slack and Microsoft Teams). For a question, a thread summary or a draft where your request carries its own context, that may be all you need. OpenAI also has a separate ChatGPT Agents app that deploys workspace agents to Slack channels (ChatGPT Agents App in Slack).
With shared team memory, the memory belongs to the agent, so a fact one person gives it can shape the answer the next person gets.
Memory is not the same as knowledge sources
Shared memory is easy to confuse with an agent's knowledge. On Runbear, knowledge sources are documents synced ahead of time from tools such as Notion, Confluence, Google Drive and SharePoint. Memory is what the agent picks up from talking with people afterward.
Put written policies and procedures in a knowledge source or in the agent's instructions, where every runtime can use them. Memory holds the small facts that come up in conversation.
How Runbear agents remember, by runtime
Every Runbear agent runs on a runtime, chosen when the agent is created. Runbear picks which runtime the recommended Standard option creates for your organization: usually Claude Agent SDK, or Anthropic for some organizations, including those with HIPAA mode. The option's description tells you which one you get (Runtimes). Each runtime remembers in its own way.
Claude Agent SDK runtime (the usual Standard choice): memory files
These agents keep memory as files in the agent's own workspace and read and update those files as they work. The Enable long-term memory toggle does not govern this memory.
Anthropic and OpenAI Responses runtimes: shared and personal facts
These agents extract short facts that point to a resource, such as which team owns a service, where a document lives, or which tool a workflow uses. They do not store tasks or opinions. Where the conversation happens decides whether a fact is shared or personal (Memory):
| 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 |
The rule follows the kind of message. A message posted in a private channel stays personal, but an @mention of the agent in that channel is shared. If someone @mentions the agent in a public channel to say the design system lives in the ui-kit repo, everyone's later answers can use that. If a user tells the agent in a direct message that they work on the billing team, only that user's later private conversations include it.
Anthropic and OpenAI Responses agents add to memory only from conversations in Slack, Microsoft Teams, Discord and the dashboard playground. HubSpot and Zendesk tickets, triggers, and requests through the REST API or the Web SDK do not add to it, so an agent that only answers tickets or runs on triggers builds no memory from that work. Turning off Enable long-term memory in the agent's settings stops them from saving and recalling facts, but the facts already saved stay. The toggle is on by default.
Classic runtime: recall of earlier messages
Classic agents recall related messages from the agent's earlier conversations. The recall covers the agent as a whole, with no split between personal and shared. Turning off Enable long-term memory turns it off.
What to check before you roll out a shared agent
- Tell your team that, on the Anthropic and OpenAI Responses runtimes, an @mention of the agent in a private Slack channel goes to shared memory.
- Check which runtime each agent runs on before you rely on how it remembers. The Standard option's description names the runtime when you create the agent.
- Keep written policies in a knowledge source or the instructions. Memory on those two runtimes holds resource facts.
- Start where many people ask the same "where is it?" questions, such as an ops or IT help channel. Memory on these runtimes holds facts like where a document lives, so that is where one person's fact helps the next person most.
Where a shared Runbear agent fits
Runbear's comparison with ChatGPT says you can keep ChatGPT for individual work and use Runbear for the agents a whole channel relies on. Those are named agents that a whole team @mentions in Slack or Microsoft Teams, such as a support agent, a sales ops agent or an HR agent (Shared AI agents for teams).
