All resources
Agents16 October 2026·6 min read

Agent memory: what to remember, what to forget, and who can see it

Memory makes agents more useful and more risky. Design what is stored, for how long and under whose permissions.

An agent that remembers a customer's preferences feels helpful. An agent that remembers everything forever, visible to anyone, is a liability. Treat memory as data with an owner, a purpose and a lifetime.

Types of memory

  • Working memory — the current task's context; discarded when the task ends.
  • Session memory — carries a conversation across steps or days.
  • Long-term memory — stable facts and preferences, stored deliberately and attributable to a source.

Design rules

  • Store facts, not transcripts — short, checkable statements are easier to correct and delete.
  • Record where each memory came from and when.
  • Scope memory by user and permission, so one person's information never appears in another's session.
  • Let people see and delete what is remembered about them.
  • Set expiry — memories not used or reconfirmed lapse.

Wrong memories are worse than none: an agent that confidently acts on stale information is hard to catch. Build correction into the design.

Related capability

This is what our Autonomous AI Agents practice is built around.

See how it's scoped