Agent Memory

What is agent memory?

Agent memory is the capability of an AI agent to retain and retrieve useful information for later decisions. It can preserve details across steps in one task, across a conversation, or across separate sessions, depending on how the system is designed.

Memory is not a single component. It is a lifecycle: decide what matters, store it, retrieve it when relevant, use it, update it, and eventually forget or delete it.

What agents may remember

Useful memory can include:

• user preferences and approved profile facts;
• summaries of earlier conversations;
• completed actions and their outcomes;
• facts learned from verified sources;
• successful or failed task strategies;
• unresolved commitments or pending work.

For example, a travel agent may remember a user’s seat preference across trips. It should not assume that a past destination or budget still applies unless that information remains relevant.

Types of agent memory

Working memory supports the current task. Short-term memory lasts for a limited conversation or session. Long-term memory remains useful across later interactions. Semantic memory stores facts and generalized knowledge, while episodic memory records specific events and outcomes. Persistent memory describes information stored so it survives beyond the current runtime.

These categories can overlap. A persistent store may contain both semantic and episodic long-term memories.

Memory versus context and state

Memory is information available for future retrieval. [Context](/glossary/context/) is what the model can use during the current inference. A memory must be retrieved into context before the model can act on it.

[Agent state](/glossary/agent-state/) tracks progress in the current execution, such as completed steps and pending approvals. Some state may be saved as memory after the task, but the concepts serve different purposes.

Model knowledge is also not agent memory. Knowledge learned during model training is embedded in model parameters and is not normally updated by one user interaction.

Memory quality and risk

Storing everything creates noise, cost, privacy risk, and retrieval errors. Memories can become stale, conflict with newer information, or preserve an incorrect inference as if it were a fact.

Reliable systems record source and time where useful, separate facts from summaries, support correction and deletion, and retrieve only information relevant to the current goal. Sensitive data should have explicit storage rules and access controls.

Why it matters

Memory gives agents continuity without forcing the full history into every prompt. It can reduce repeated questions and support longer tasks, but only when writing and retrieval are selective and trustworthy. See [Memory in AI Agents](/memory-in-ai-agents/) for the full lifecycle and its relationship to context, RAG, state, and model knowledge.