Skip to main content
Fleet RLM uses one local User and one Workspace. The Workspace owns everything durable: Sessions, files, memory, Skills, Attachments, and Artifacts.

Domain model

Persistence

Fleet stores Sessions, Runs, and Turns in the database named by FLEET_DATABASE_URL.
  • Use SQLite for local development. The example .env sets sqlite+aiosqlite:///./.fleet_rlm/local.sqlite3.
  • Use PostgreSQL at the Alembic head for durable deployments. See Deploy with PostgreSQL.
Fleet never migrates the database on startup. Run uv run python scripts/database.py upgrade after install and after every upgrade.

The Workspace Volume

One Daytona Volume holds each Workspace under workspaces/<workspace_id>. Fleet mounts it two ways. Model code reaches the wider Workspace only through host tools. See Sandbox tools.

Attachments and Artifacts

  • Upload an Attachment with /attach <path> in the terminal or POST /api/attachments. Pass up to 32 Attachment IDs in the next Turn request.
  • Fleet creates Artifacts with the create_artifact and publish_workspace_artifact tools. An Artifact stays a private candidate until Turn Commit. There is no public endpoint to create Artifacts.
  • storage.max_upload_bytes and storage.max_artifact_bytes cap both sizes. Each defaults to 10 MiB.

Workspace memory

Workspace memory is file state shared across Sessions. Fleet reads it with read_workspace_memory, list_memories, and search_memories, and writes it with remember, edit_memory, update_workspace_memory, and forget. rlm.autonomous_memory_categories lists the categories Fleet may propose on its own through propose_memory. The shipped configuration leaves it empty. Memory is not a substitute for Turn Commit. Only committed Turns form the conversation history.

Cancellation

Cancel a Run with /cancel or PUT /api/runs/{run_id}/cancellation. The stream emits one turn_cancelled event. After settlement, history shows the original input and an assistant message with a cancelled status and the text Turn cancelled. Tombstones never contain reasoning, code, output, or tool evidence.
Last modified on October 9, 2026