Memory that compounds

Every Space, every correction, every run makes the next one sharper.

v2.1

release · May 27, 2026

  • Preferences: corrections become standing rules
  • Procedures: what works gets reinforced per task
  • Knowledge graph: linked, not siloed
  • Every memory is tied to a source you can edit

Memory is the difference between an agent that helps once and one that gets better. This release turns three kinds of memory on across every Space.

  • Preferences. Tell an agent "always cc the account owner" once. It doesn't need to be told again.
  • Procedures. Agents track outcomes per task. The hundred-and-first run is measurably better than the first.
  • Knowledge graph. People, projects, tools, and decisions are linked, not siloed. What the org learns in one corner is reachable from every other.

Every piece of memory is tied to a source. Open it, see where it came from, edit it, or remove it.