# Memory
> What an agent keeps between conversations, how it decides, and how to inspect, correct, or delete what it stored.
**Memory** is what an agent carries between [conversations](/docs/work/conversations/) — facts about you, your preferences, how you like work done. It is stored deliberately and you can read every entry.
## When to use it
Memory is for facts with a long shelf life that you should not have to repeat: which timezone you work in, who your team is, that you want numbers rounded, that "the report" means the one in a specific folder.
It is not a filing cabinet. Content belongs in an [artifact](/docs/capabilities/artifacts/); background for one body of work belongs in a [project](/docs/capabilities/projects/).
## How an agent uses it
The agent decides what is worth keeping and writes it down — often after you correct it, or state a preference. At the start of a conversation it loads an index of what it knows, and pulls in the entries relevant to the task rather than everything it has ever stored.
You can also just tell it: *"remember that I never want Friday meetings"* is a normal instruction and it will store it.
## Inspecting and correcting
Memory is listed in [Agent Builder](/docs/build/agent-builder/), under the agent's Identity section. Every entry is readable, and you can delete anything that should not have been kept.
Do check it occasionally. Agents write down what they infer, and an inference stated as fact — *"prefers short replies"* after one impatient message — will quietly shape answers for weeks.
## Turning it off
Memory is a [built-in tool group](/docs/capabilities/built-in-tools/). Switching it off for an agent stops it storing and recalling, and stops the memory index being loaded at the start of a conversation. Existing entries are kept; the agent stops using them.
This is the right setting for an agent that handles other people's data, or one you want stateless and predictable.
## Limits
- Memory is per agent, not per account.
- The agent decides what to store; you cannot pre-write a list of memories, but you can tell it what to remember and delete what it got wrong.
- Deleting an entry stops it being recalled in future runs. It does not edit conversations where it was already used.
- Memory is not searchable content storage. Large documents belong in artifacts.
## Related