Skip to content

Managing Memory

Agents remember things about you and their own context across conversations. This guide covers how memory works from your perspective and how to manage it.

There are two main types of facts agents store:

When you share personal information, preferences, or context, the agent stores it as a user fact. All agents can access these facts.

Examples:

  • “I prefer Python over JavaScript”
  • “My timezone is PST”
  • “I’m working on a project called Atlas”

Each agent can store private notes about its own context. Other agents can’t see these.

Examples:

  • “Last research topic was cloud cost optimization”
  • “User’s project Atlas uses PostgreSQL and Redis”

Just say it naturally:

  • “Remember that I prefer dark mode”
  • “Keep in mind that our team uses Slack for communication”
  • “Note that the deadline is March 15th”

The agent stores the fact and confirms what it remembered.

If the agent has outdated information, tell it:

  • “Actually, we switched from Slack to Discord”
  • “Update my timezone. I moved to EST”

The agent overwrites the old fact with the new one.

  • There’s no UI to browse all stored memories directly
  • You can’t selectively delete individual facts from the chat interface
  • Memory compaction happens automatically. You can’t trigger it manually

The platform automatically curates memory over time:

  • Duplicate facts are merged
  • Contradictory facts are resolved (newer wins)
  • Stale information is archived

This runs in the background every few hours. You don’t need to worry about memory growing too large.

  • Be explicit when you want something remembered long-term
  • If an agent seems to have forgotten something, it might have been compacted. Just remind it
  • Different agents have different memories, so context shared with one agent isn’t automatically known by another (though user facts are shared)