Skip to main content
Every agent has a workspace in its sandbox, shared by all its sessions, and a memory the harness keeps across conversations. Both live on the sandbox’s disk and survive sleep. You can read and write them from your backend at any time.

The workspace

The workspace root is /home/user/workspace in the sandbox. File paths in the API are relative to it, and so is each session’s cwd: a session created with cwd: "repos/web" runs its harness in /home/user/workspace/repos/web.

Memory

Memory is what the harness keeps across conversations: text files the agent writes as it works and reads at the start of each conversation. Each session’s instruction file (CLAUDE.md for Claude Code, AGENTS.md for Codex) names the memory directory and includes its MEMORY.md, up to 32 KiB; the harness reads any other memory file itself when it needs it. Memory is the agent’s data, never configuration: a new agent starts with none.
Memory paths are relative to the agent’s memory directory. memory.list(), read, write, and delete work like their file counterparts, on text.

Rules