Feature
Spaces: projects with their own instructions, agents and memory
A Space in Milly Lab groups the chats of one project or client and gives every new chat there the same starting point: a system prompt, a default model, an optional agent that answers, and the Space's own memory. Around Spaces, the Console adds agents, memory collections, connected accounts, skills, MCP servers and schedules.
Last updated:
What does a Space hold?
A Space holds its chats and the defaults for every new chat in it: a default model, a system prompt, an agent that answers or none, and the settings of the Space's memory.
| Setting | What it does |
|---|---|
| Default model | New chats in the Space start with this model. If your plan does not include it, the chat keeps the model you picked and tells you so. |
| System prompt | Instructions added to every chat started in the Space, such as a client's tone of voice or a project's rules. |
| Answers as | An agent that answers every chat in the Space, or plain chat without an agent. |
| Distill this Space | Builds a memory collection from the Space's chats. |
| Use the Space's memory | Adds that collection to chats opened in the Space. Off by default. |
| Sealed memory | A two-way wall: chats in the Space use only its own memory, and facts learned there never appear in other chats. |
| Make an agent | Drafts an agent from the Space's instructions, model and collections. |
| Export to Google Docs | Writes the Space's chats to Google Docs in a Drive folder you choose. Needs a connected Google account. |
A chat belongs to a Space only when you put it there. Deleting a Space removes the Space and its settings; the chats inside it are kept.
What is an agent in Milly Lab?
An agent is a saved assistant with its own instructions, model, tools and knowledge. You call it with @handle in any chat, make it the agent that answers in a Space, or run it on a schedule.
- Create one in the Console, or let Milly Lab draft it from up to 20 of your chats or from a Space. Nothing becomes an agent until you review the draft.
- An agent uses only what you tick for it: web search, running code in a sandbox, making documents, utilities, connected accounts, MCP servers, skills, memory and memory collections.
- Approvals still apply: anything that writes, sends or deletes asks you first, as in a normal chat, unless that tool is set to Allowed.
- Changes save as a private draft while you edit, and a Test pane talks to the draft. Save makes the version that chats use.
- Publish submits an agent to the public gallery, where a moderator reviews it first. Connected accounts, tokens, MCP keys, memory facts and chats are never published.
- One account can hold up to 200 agents.
How does memory work?
Memory is a paid-plan feature that stays off until you switch it on. Milly Lab then keeps short facts about you, summaries of past chats and named memory collections, and every answer that used memory shows a Used memory chip.
| Plan | Memory per message, characters | Memory collections |
|---|---|---|
| Free | Not included | Not included |
| Starter | 8,000 | 1 |
| Pro | 10,000 | 2 |
| Max | 18,000 | 10 |
| Advanced | 28,000 | 25 |
| Enterprise | 28,000 | 25 |
- Facts come in four kinds: about me, preference, project and rule. You can read, pin and delete each one, or forget everything at once.
- A memory collection is distilled from up to 20 chats or from an imported file: Word, Excel, PowerPoint, text, Markdown, CSV, JSON, XML, HTML, or a ChatGPT or Claude data export.
- A collection can become a notebook that answers questions with sources and can record a short audio overview.
- Recalled memory is sent to the model as input, so it counts toward your usage, within the per-message limit in the table.
Which tools can chats and agents use?
Chats and agents can use connected accounts, skills, MCP servers and plugins. Reads run freely; writes follow each account's setting, which asks you first by default.
| Tool | What it is | Limit |
|---|---|---|
| Connected accounts | Mail, calendars, documents, team chat, CRMs, marketplaces and accounting services, such as Google Workspace, Microsoft 365, Slack and Notion. Writes are set to Allowed, Ask before writes (the default) or Off. | Actions that move money or stock always ask |
| Skills | Instruction sets (SKILL.md files) the model loads only when a request needs them. Built-in skills include business letters in Tajik and Russian, contract review for CIS contracts and meeting notes. | Up to 100 of your own |
| MCP servers | Any remote MCP server, added by its https address, with sign-in (OAuth), a bearer token or custom headers. | Up to 25 servers |
| Plugins | Bundles of skills, agents and MCP settings in the Claude Code, Agent Plugins (OpenAI, Codex) or Gemini CLI formats. | Up to 10 plugins per upload |
How do schedules work?
A schedule runs a prompt on its own, as a plain chat, inside a Space or as an agent: as often as every 15 minutes, hourly, daily, on weekdays, weekly, monthly, once or on a custom cron. Each run is an ordinary chat turn and uses your normal allowance.
- Up to 50 schedules per account, and up to 20 new ones in 24 hours.
- Each run opens a new chat or is added to one conversation you choose.
- Results always land in the Tasks inbox and can also arrive by e-mail, push notification or a signed webhook.
- While you are away, each tool that would ask for approval in a chat is either drafted for your approval or allowed in advance. Anything that moves money or stock never runs unattended.
- After several failed runs in a row, a schedule pauses itself.
Which plans include Spaces?
Spaces, agents, skills, MCP servers and schedules are available on every plan, including Free; memory and memory collections need a paid plan. What an agent or a schedule can do still depends on the models your plan includes.
Compare plans and the models they include