Agent memory glossary
Plain definitions of the terms behind AI agent memory: experience and procedural memory, playbooks, memory poisoning, holdouts, plan replay and more.
- Experience memoryExperience 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.
- Procedural memoryProcedural memory is an agent's stored knowledge of how to do a task, as opposed to facts or past events. How it is learned from verified runs.
- PlaybookAn agent playbook is a reusable procedure for one kind of task, with steps, conditions and pitfalls, given to the agent before a similar task.
- Failure signatureA failure signature identifies one kind of agent failure by tool, error and input shape, with its frequency and the fix that resolved it.
- Memory poisoningMemory poisoning plants instructions in an AI agent's long-term memory through content it processes, so the agent obeys them in later tasks.
- Indirect prompt injectionIndirect prompt injection hides instructions in content an AI agent processes, such as web pages, emails or tool output. Examples and mitigations.
- Taint trackingTaint tracking labels agent inputs as trusted or untrusted and follows that label through a run, so untrusted input cannot change agent memory.
- Holdout groupA holdout group is a random share of traffic left without a feature, such as agent memory, so its real effect can be measured against the rest.
- Plan replayPlan replay lets an agent reuse a proven plan instead of planning from scratch, when a task closely matches one it has solved reliably before.
- Circuit breakerA circuit breaker stops an AI agent from retrying a tool call that cannot succeed, ending retry loops before they burn tokens and time.
- Agent regression testingRegression testing for AI agents replays known tasks after a change and checks outcomes and tool calls, so prompt or model changes cannot break them.
- LLM-as-a-judgeLLM-as-a-judge uses a language model to evaluate an agent's output when no automatic check exists. Its strengths, weaknesses and safe use.
- Model migrationModel migration is switching an AI agent to a new language model while keeping its behaviour on real tasks. Risks and a safe process.