What is Agent memory?

There is no single standard for agent memory, so this entry describes one concrete design: Anthropic's memory tool for the Claude API.

Anthropic's memory tool gives Claude a directory of files it can write to and read from, so what it learns in one conversation can be picked up in a later one. Claude can create, view, edit, and delete those files. The docs describe this as just-in-time retrieval: instead of loading everything up front, the agent notes what it learns in files and reads them when needed, so the context window stays on the task at hand.

The tool runs client-side. Claude only requests file operations under a /memories path, and the developer's application carries them out against storage it controls, such as a per-user directory or database keys. The commands are view, create, str_replace, insert, delete and rename. When the tool is enabled, the API adds instructions telling Claude to check its memory directory first and to assume its context may be reset at any moment.

The page puts safety work on the developer. Handlers must reject paths outside /memories to prevent directory traversal. It also advises capping file sizes, expiring old files and stripping sensitive data before anything is written.