Memory

An agent that starts from zero every time is slow and expensive. Helios agents carry context forward in two ways.

The persistent workspace

Files an agent writes under ~/workspace survive between runs. For a workflow, every run shares the same workspace, so an agent can leave itself notes, cached data, or state:

Keep a file at ~/workspace/processed.json listing the issue numbers you've
already handled. Skip any issue already in that list.

This is the memory you control directly. See sandboxes.

API memory

When an agent works out how to accomplish a task against an integration's API — which endpoints to search for, which calls to make in what order — Helios records that procedure. The next time an agent takes on a similar task against the same integration, it starts with that record instead of rediscovering the API.

This memory is:

• Scoped to your organization and the specific integration. It is never shared across organizations.
• Written automatically after a task completes successfully, and read automatically before a similar one begins.
• Procedural, not data. It records how an API was called, not the contents of your records.

There is nothing to configure and nothing to manage. It makes repeat tasks faster and more reliable.

What is not memory

A chat's message history is part of that chat, not shared with other chats or with workflows. Forking a chat copies its history and workspace into a new one; that is the only way conversation context moves between sessions.