What is experience memory for AI agents?

Experience memory is an agent's long-term memory of what worked and what failed on its own tasks, recalled before similar tasks so it improves with use.

Updated

Experience memory is long-term memory for an AI agent made from its own past runs: procedures that worked, failures it hit and the fixes that resolved them, recalled before similar tasks so the agent improves with use.

How it differs from other agent memory

Agent memory comes in several kinds, and they answer different questions:

KindAnswersExample
User or semantic memoryWhat is true about this user or topic?"The customer prefers email"
Conversation memoryWhat was said earlier?The last ten messages
Experience memoryWhat worked, and what failed, for this kind of task?"Cancellations need an uppercase reason code; if it is rejected, use one of the codes the error lists"

User memory personalises an assistant. Experience memory makes an agent better at its job.

What makes experience memory work

  • Verification. Only runs confirmed to have worked should teach, or the agent memorises its own mistakes. See LLM-as-a-judge.
  • Gated recall. A procedure for a different kind of task is worse than none, so memory must be recalled only when it clearly applies.
  • Protection. Memory learned from untrusted input is an attack surface. See memory poisoning.
  • Measurement. A holdout group shows whether memory actually helps.

In Retrack

Retrack is experience memory as a service: playbooks for what worked, failure signatures for what failed, a memory firewall and a built-in holdout. See how it works, or the agent memory guide for every kind of agent memory and the tools compared.