What Is Dust? AI Agents on Company Knowledge
Dust is a platform where teams build AI agents that answer from company data and act in company tools. Admins connect sources such as Slack and Notion. Builders then give each agent instructions plus the knowledge and tools it needs, on a frontier model of their choice. Since August 2025, agents can also keep a per-user memory across conversations. This guide explains how those pieces fit together and what Dust costs. It also shows which layer lets an agent learn from use: agent memory, scoped to one user.
What is Dust, in one sentence?
Dust is a workspace where teams build and share AI agents that work on the company's synced data and connected tools, using frontier models from several labs.
The Dust docs describe a platform for "customizable and secure AI agents" that integrate with company data sources. Users call agents in a chat by mentioning them with @. A default @dust agent searches all the internal data the workspace has synced. Teams add their own specialized agents, and several agents can work in one conversation.
How do Dust's knowledge, tools and memory differ?
An agent in Dust can draw on three kinds of context. They differ in who controls them and how long they last.
| Company data (connections) | Tools | Agent memory | |
|---|---|---|---|
| What it holds | Synced documents from Notion, Slack, Drive, GitHub and others | Live access to apps such as Salesforce, HubSpot or GitHub | Facts about one user, learned in conversations |
| Who sets it up | Admins choose what syncs | Admins configure, builders add to agents | Builders add the Memory tool. The agent decides what to save |
| Scope | Company-wide or per space | Per agent, with personal or workspace credentials | Private to one user and one agent |
| How long it lasts | Kept in sync with the source | Each call reads live data | Until the user deletes it |
| Can you inspect it? | Yes, in the source app | Yes, in the source app | Yes, in the agent's Memory tab |
Sources: Dust docs on data sources and tools. The memory column follows the agent memory page.
How does Dust connect agents to company data?
Data enters Dust as data sources. The knowledge overview starts with connections that sync automatically. MCP-based tools reach third-party systems, and a crawler ingests public websites. Static folders and uploaded files count too, as do custom connections built on the API.
Admins pick exactly what syncs, even a single page or channel. The data access FAQ spells out the limits. Slack attachments and messages from other bots are not indexed. Dust reads Google Docs and Slides from Drive, plus small text files. PDFs are an opt-in. For GitHub, Dust syncs issues, pull requests and discussions, but not repository code. Images and comments inside Google documents or Notion pages are skipped.
Agents reach that data through retrieval tools such as search. Default tools add web search and a computer that runs code. Agents can also create images or call another agent. App tools go further and write to the source, for example by creating a Notion page or a GitHub pull request.
How does Dust agent memory work?
Memory is a tool that a builder adds to an agent. Dust announced it on August 14, 2025, and made it opt-in on purpose. Its example is a code review agent that should apply the same standards to everyone. Memory could bend those reviews toward each developer's habits.
Once the tool is on, the memory docs say the agent decides what is worth remembering and saves it during conversations. Users can also ask it to remember something. The agent retrieves memories when it judges them relevant. Each user can browse and delete their memories in the agent's Memory tab, and deletion is permanent.
Memories are private to each user. Each agent also keeps its own memory space, so a preference you teach your sales agent stays with that agent. Dust's blog says memory data is encrypted at rest and never used for model training.
What does Dust cost?
Dust prices by seat and by credit. The pricing page lists a Business plan for teams of up to 100 people with three seat types. Free seats get 500 credits for life. Pro seats cost 24 euros a month billed yearly (30 euros monthly) for 8,000 credits. Max seats cost 120 euros a month yearly (150 euros monthly) for 40,000 credits. Prices exclude VAT, and unused credits do not roll over.
Credits are charged per message and depend on the model and the tools used. Business includes 20+ models and SSO. Its plan comparison allows up to 3 of the 20+ data source connectors. Enterprise is custom priced and lifts that cap, adding SCIM as well. It also offers audit logs and single-tenant deployment.
What should you read in the fine print?
Dust's security page claims SOC 2 Type II certification and GDPR compliance, and says it "enables HIPAA compliance." It also says customer data is never used to train models and that third-party model providers keep no data. These are Dust's own statements, and the Trust Center is where an auditor would check them.
Agents see only what admins synced. A page left out of the sync, or synced late, never reaches an answer. Code in GitHub repositories is not indexed. Memory is per user and per agent, so one person's corrections do not help a teammate. Dust's blog describes cross-user learning as a goal it is working toward.
The bottom line
Dust fits teams that want shared agents on their own documents and apps without building retrieval themselves. Start with a Pro seat for builders and free seats for occasional users, and watch credit use on tool-heavy agents. Turn on memory only for agents where personal context helps, such as sales or support. Keep shared rules in agent instructions or a synced document, where everyone gets them.
How do you share agent memory across tools and teammates?
Dust keeps memory private to one user and one agent, a sensible default for a company chat platform. A fact your support agent learned is invisible to your research agent. None of it reaches the coding agents your engineers run in Claude Code or Codex.
memU gives coding agents one memory that several hosts share. A scheduled background task slices new session logs into jobs. The agent distills them into memory and skill Markdown, which you can read and edit. memU makes no LLM calls itself. It stores and embeds what the agent prepared, then retrieves it on request. A standing instruction in CLAUDE.md or AGENTS.md tells the agent to retrieve before answering. Supported hosts, including Claude Code and Codex, share one backend. A lesson learned in one is available in the other. memU does not connect to Dust.
The hosted version is free and syncs across devices. Get an API key at memu.so and send your agent one message:
Read https://memu.pro/SKILL.md, follow its instructions to install and configure memU, API Key is
YOUR_MEMU_API_KEY.
The self-hosted path and the host support matrix are in the memU GitHub repository.
FAQ
What is Dust used for?
Teams use Dust to build AI agents that answer from synced company data and act in tools like Notion and GitHub. Dust also offers general model access through agents like @gpt.
Does Dust have agent memory?
Yes. Builders can add the Memory tool to an agent. The agent then saves facts about each user across conversations. Users can review and delete them in the agent's Memory tab.
How much does Dust cost?
The Business plan has free seats with 500 lifetime credits. Pro seats cost 24 euros a month billed yearly, and Max seats cost 120 euros. Enterprise pricing is custom.
Can Dust read my code on GitHub?
Not through its GitHub connection. The GitHub connection skips repository code. It syncs issue and pull request threads, plus discussions. The GitHub tool can still search and open pull requests.
Does Dust train models on company data?
Dust says no. Its security page states that customer data is never used to train models and that third-party model providers retain no data.