loading
loading
A client-side tool that lets the model read and write files in a persistent directory so it can carry knowledge across separate sessions.
Declared as {type: "memory_20250818", name: "memory"}, it gives the model commands (view, create, str_replace, insert, delete, rename) over a /memories directory whose storage backend you implement. Where context editing and compaction manage state within a single session, memory persists across sessions and survives process restarts — an agent can jot down lessons in one run and consult them in the next. Never store secrets or unguarded PII in memory files, and in multi-user systems scope the directory per user. Newer models are notably better at using a memory surface, so it's worth telling the agent where it is and giving it a note-taking format.
Plainly
Think of Memory Tool as a tool from the builder toolbox. A client-side tool that lets the model read and write files in a persistent directory so it can carry knowledge across separate sessions.
In practice
Use it when you are installing, editing, testing, reviewing, or shipping code. In practice, define the owner, input, output, and failure mode before you rely on it.