.memory.md and live in the Space’s memory/ folder.
What a memory file looks like
A memory file is plain Markdown with a small frontmatter block. You write the starting content; the agent appends to it as it runs.Attaching memory to an agent
On the Agent canvas, bind memory files to the agent through its read and write sets:- Read set — one or more memory files the agent can see. Their contents are injected into the run before the agent starts thinking.
- Write set — at most one memory file the agent can add to, using the built-in
remembertool.
An agent with no write set can still read. Memory that only ever gets read is a
perfectly good way to hand an agent standing context — house style, account history,
naming conventions — without letting it edit that context.
Seed memory and runtime memory
There are two kinds, and the difference matters when you deploy. Seed memory is what you author. It ships with the agent, and every deployment starts from it. Runtime memory is what the agent writes while running. It stays with the instance that wrote it and never travels back into your Space. So a deployed agent begins with the memory you curated, accumulates its own on top, and your authored file stays exactly as you left it.Watching an agent use its memory
When an agent runs, the trace shows memory being read at the start and eachremember
call as it happens, so you can see what it chose to keep. If a memory file can’t be read,
the run says so rather than continuing silently without it.
Compose an agent
Bind read and write sets on the agent canvas.