Your bot: workspace, persona & memory

Your bot is a Claude agent whose identity, instructions, knowledge, and standing rules are assembled by Agent Composer. Each session starts with a fresh model context, so GATE builds the bot again from durable configuration and memory before the conversation continues.

The workspace still exists for working files, downloads, daily notes, and project material. It is no longer the source of truth for the bot's persona and boot rules.

Agent Composer

Composer stores the bot's configuration as small, reviewable blocks rather than one ever-growing Markdown file. A block has a title, body, kind, scope, audience, enabled state, and version history.

The common block kinds are:

• Persona — who the bot is and how it sounds.
• Knowledge — stable facts and background it should know.
• Instruction — standing directions for how to work.
• Rule — a constraint the bot should keep following.
• Memory — a generated index into durable memory.
• Runbook — an approved how-to the bot discovers by summary and opens only when needed.

Before a chat turn, GATE assembles the applicable blocks from broadest to most specific: platform rules, house rules, organization blocks, the bot's own blocks, and finally the current human's personal instructions. That last user band reaches only that person's conversations; it is never shared into another user's session.

Each block also chooses where it applies: chat, code sessions, or both.

What the workspace still does

The bot keeps a real server workspace for the files it creates and consumes: uploads, downloads, source checkouts, scratch material, daily notes, and other job artifacts. You can browse those files from the bot's dashboard.

MEMORY.md is a compatibility fallback. Once the bot has a generated Composer memory-index block, that block replaces the file at boot; until then, the bot may still read MEMORY.md so it does not lose its existing memory during the transition.

Durable lessons

When a bot learns a lasting lesson—especially after a correction—it files a bot-level rule block on its own Composer board. The lesson is live on the next assembled prompt, deduplicated by title, and visible to humans with history and an enable/disable control.

That is different from:

• a one-off task note, which belongs with the task,
• a fact to remember, which belongs in memory, or
• an organization-wide rule, which needs the governed promotion path.

A self-filed lesson applies only to that bot. It cannot write a rule onto another bot.

Picking up edits mid-session

An already-running session does not automatically inherit a configuration edit that landed after it started. Ask the bot to reload. GATE reassembles the Composer document for that specific session and preserves the conversation history.

If assembly fails, the bot receives a loud configuration-error block instead of quietly falling back to stale workspace files. That makes a broken composition visible rather than giving you a bot that looks healthy but is running the wrong rules.

Going deeper

For how bots fit into teams and accounts, see Organizations, users & bots; for how conversations work, see Sessions; and for how your bot acts on your behalf in other services, see Credentials & delegation.