Using Memory
You do not need to configure anything to start using memory. The agent automatically recalls relevant memories and stores new ones as you chat. This page covers how to actively shape what the agent remembers.
Teaching the agent
The most direct way to create a memory is to tell the agent to remember something:
"Remember that I like concise answers with code examples."
This becomes a preference in your personal memory for this workspace (private to you; it does not follow you into other workspaces).
Personal is for how you work here. Workspace is for shared project knowledge. Prefer telling the agent to save team facts for the workspace:
"Remember that our API uses snake_case for all response fields — save that for the workspace."
To share a memory with your whole team or organization, ask explicitly:
"Save this for the whole workspace: our API uses snake_case for all response fields."
"Save this as an org-wide policy: all customer-facing text must be reviewed by the content team before launch."
The first becomes a workspace project_fact; the second an org_policy in organization memory.
Correcting the agent
When the agent gets something wrong, correct it directly:
Agent: "I'll set up the project with npm..."
You: "We use pnpm, not npm."
The agent saves a correction to your personal memory in this workspace. In future conversations here, it will use pnpm without being told again. If you want the whole team to inherit the fix, tell the agent to share it ("save that for the workspace").
"That's wrong — I prefer dark mode code blocks, not light."
What the agent remembers automatically
Beyond explicit instructions, the agent may store memory items when it identifies reusable information during a conversation. For example:
- You describe a deployment process step by step — the agent may save it as a procedure.
- You and the agent agree on an approach — it may store the outcome as a decision.
- You share a fact the agent did not know — it may record it as a project_fact.
Each automatically created memory includes a confidence score. Items from explicit instructions ("remember that...") typically get higher confidence than items the agent infers from context.
Memory in action
Here is a practical example of how memory accumulates and helps over time:
-
Day 1 — You tell the agent: "Save this for the workspace: our backend is FastAPI with PostgreSQL, deployed on AWS ECS." The agent saves a workspace-scoped project_fact so the whole team inherits it.
-
Day 2 — A teammate asks the agent to help write a database migration. The agent already knows the stack is PostgreSQL and uses the correct syntax without asking.
-
Day 3 — You correct the agent: "We use Alembic for migrations, not raw SQL — save that for the workspace too." The agent saves a correction to workspace memory.
-
Day 4 — Another teammate asks for help with a new migration. The agent recalls both the PostgreSQL fact and the Alembic correction, and produces an Alembic migration file.
Scope selection
The agent chooses scope from context and your instructions. Personal stays in the current workspace only. Shared project knowledge should use workspace.
| Content pattern | Typical scope | Typical type |
|---|---|---|
| "I prefer..." / "I like..." | Personal | preference |
| Correcting how the agent works with you | Personal | correction |
| "Our project uses..." / "The codebase..." | Workspace | project_fact |
| Agreeing on a team approach | Workspace | decision |
| Describing a shared deploy/process | Workspace | procedure |
| "Save this for the workspace: ..." | Workspace | project_fact / procedure / decision |
| "Save this as an org-wide policy: ..." | Organization | org_policy |
To promote a memory beyond yourself, say so when you ask the agent to remember it ("save that for the whole workspace"). You cannot change the scope of an existing item from the Memory Center, so be explicit up front.
Tips
- Be specific. "Remember that we use 4-space indentation in Python files" is more useful than "remember our coding style."
- Correct early. The sooner you fix a mistake, the less likely the agent repeats it in other conversations.
- Check the Memory Center periodically. Over time, memory items may become outdated. Archiving stale items keeps the agent's context clean.
Next steps
- Managing Memory — Review and archive memory items from the Memory Center.