Where your team and its agents share the same doc.
Dispatch, propose, approve, learn — in one workspace.
Contextful turns the document into the control plane. Write a command in plain English, get a structured proposal back, quote the underlying data inline, and approve before anything ships. Every accepted edit teaches the agent.
Four primitives that make humans and agents collaborate
Commands inside rich text
Type /agent anywhere in a doc. Mention people, agents, or data sources with @. The command lives next to the context that produced it — no Slack DM, no forgotten thread, no copy-paste of an SQL result.
- Slash commands route to any agent on the platform
- Mentions resolve to people, agents, files, or queries
- Threads attach to the line that triggered them
Proposals, not autopilot
Agents return structured proposals — diffs, drafts, transactions, write-backs — that humans review before anything reaches a real system. Approve, edit, reject, or counter. The agent treats your edit as the new ground truth.
- Diff view for every write-back to QuickBooks, Slack, CRM, etc.
- One-click approve, edit-and-approve, or reject with reason
- Audit trail captures who approved what, when, and why
Quote and visualize live data inline
Pull a row, a chart, a query, or a file snippet directly into the document. Quotes stay live — re-run on demand, with provenance baked into the citation. Reviewers see the same evidence the agent saw.
- Inline tables, charts, and code blocks bound to live queries
- Provenance metadata: source, fetch time, capability token
- Snapshot or live mode — pin to a moment, or always re-run
Agents with self-improving memory
Every approved edit, rejected proposal, and inline correction becomes a memory. Agents revise their own playbooks — what to escalate, what wording your team prefers, which signals matter — without retraining a model.
- Memories scoped per workspace, per team, per role
- Inspectable, editable, and exportable like any other data
- Versioned — roll back a bad lesson the same day
One thread, end to end
How a single request moves through dispatch, proposal, citation, approval, and memory — without leaving the document.
Dispatch
An analyst writes /close-month at the top of the close doc and @-mentions the QuickBooks workspace. The command is a structured object on the doc, not a chat message.
Propose
The agent returns a proposal: open items, suspected accruals, draft journal entries. Each line is quoted with the underlying ledger row, fetched through a sandboxed connector under the analyst's role.
Approve
The controller reviews the diff, edits two amounts, rejects one entry with a one-line reason, and approves the rest. Only approved entries write back to QuickBooks.
Learn
The edits and rejection become memories. Next month, the agent flags the same pattern earlier, drafts the controller's preferred wording, and skips the entry it learned not to propose.
What a self-improving memory actually is
Memory on Contextful is data, not a black box. You can read it, edit it, version it, export it.
Scoped
Memories live at four levels: agent, workspace, team, and org. A finance agent's lessons don't bleed into HR. A team's preferences don't override an individual's edits.
Inspectable
Every memory has a provenance trail — which interaction produced it, who approved or rejected the underlying proposal, which document it lives next to. No mystery weights.
Editable
Memories are Markdown with structured frontmatter. Open one, edit it, delete it, or write a new one by hand. Agents read the same store you do.
Versioned
A bad lesson is a one-click rollback, not a six-week retraining cycle. Memories are diffable, auditable, and exportable as plain files.
Built so collaboration doesn't break governance
Permissions follow the human
An agent invoked from a doc inherits the dispatcher's role through capability tokens. Reviewers see only what they're allowed to see — even when an agent did the fetching.
Approvals are first-class
Every write-back to a connected system is gated by an approval. Approval policies are configurable per workspace — single approver, dual approver, threshold-based, or fully autonomous for low-risk actions.
Audit by default
Dispatch, proposal, citation, approval, and memory update are each their own audit event. Replay any thread end-to-end — for a customer, an auditor, or an incident review.
Bring your team and its agents into the same doc.
Get early access, or read how the collaboration model works under the hood.