Agent memory: what agents recall between runs

A digital worker doesn't start cold every time. Agent memory lets it recall its own similar past runs and use them as context for a new one — so recurring cases are handled consistently rather than reinvented.

How it works

When a run starts, the agent looks back over its prior runs for the closest matches and includes a short summary of how those were handled. It's scoped strictly to that agent within its workspace — memory never crosses tenants.

What it's good for

Repetitive, patterned work: the same kinds of refund, the same invoice exceptions, the same support questions. Memory nudges the agent toward the outcome that worked last time. It complements — it doesn't replace — the guardrails and approval gates, which still apply to every run regardless of what memory suggests.