Skip to main content
Pro and Team workspaces can turn their table records into connected knowledge for people, workflows and AI agents. Open Workspace Brain in the sidebar. Existing and new table records are included by default, with the same IDs and revisions. Webhooks and data sources contribute records through the table’s Inputs mappings in Data workspace.

Explore your connected data

Graph starts with a workspace overview. Each group represents an included table or standalone knowledge. Counts cover the entire workspace, even when it has more records than a single page can display. Dashed contains connections show membership in the Brain. Other connections summarize saved relationships; a loop represents relationships between records in the same group. Relationships are created explicitly by people, workflows or agents. Select a group or use Find a record to browse its records. Select a record, then choose Explore connections to follow its relationships across tables and record pages. Use Relationship and Direction to narrow the view. Select a neighbor and explore again to continue through the graph. Workspace overview returns to the map, and Arrange restores an automatic layout. Record pages show up to 100 records. Connection exploration shows up to 100 nodes and 200 edges within 64 KiB; notices identify truncated views. The overview shows up to 500 relationship groups while its total still counts all included saved relationships. Complete properties load when you select an individual record.

Choose what belongs in your Brain

Choose Manage included data to include or exclude whole tables. An excluded table contributes neither existing nor future records to Brain discovery and agent context. To omit a single record, select it in Graph or Grid and choose Exclude from Brain. Excluded records and their saved connections remain stored. Table records still work in their original tables, views, dashboards and table workflows. Including a table again restores eligible records and preserves your individual exclusions. To restore a record, open Grid, choose Brain inclusion → Excluded, select the record, then choose Include in Brain. Its table must also be included. The excluded-record count is a shortcut to that filter. Writers can change these choices; readers can inspect them. Inclusion changes have their own revision, so ordinary record edits do not overwrite them. If someone else changes inclusion first, refresh and review before saving again.
Inclusion controls Brain discovery and context retrieval, including every hop of a graph workflow or agent query. Excluded records cannot bridge other records. It does not change existing table permissions or prevent authorized curation queries from reading excluded data. Context already sent to an agent stays with that agent.

Edit and connect knowledge

Grid gives you a spreadsheet. Choose a table to edit its typed fields, or choose Entire workspace to inspect names, types, JSON properties and revisions. The Brain inclusion filter lets you review included, excluded or all records. Double-click a cell to edit it. Select a record and choose Edit node for the full editor. Table-backed edits also update the original table and its dashboards. Use New node for standalone knowledge such as a policy, customer briefing or agent memory. Give it a name, a type such as knowledge, and JSON properties. Use Connect nodes to create a directional relationship such as informs or depends_on. The Graph view lets you pan, zoom, move and inspect nodes and their connections. Node positions are local to the current view. Choose Expand canvas in Graph or Execution to use the screen for exploration. Click a node to open its details alongside the expanded canvas. Fit graph to view brings the visible graph into view; drag the background to pan and use Ctrl/Cmd + scroll to zoom. Press Escape or Exit expanded view to return. Use Browse data to open the table and dashboard navigation when you need it. Hide sidebar gives the Brain the full workspace width again. Table-backed nodes keep the table’s field schema and derived name. Execution trace nodes are read-only. Deleting a table-backed node also deletes its table record; deleting a connection leaves both records intact.

Preview agent context

Select a node and choose Preview agent context. The preview expands two hops, with up to 40 nodes, 80 relationships and 16 KiB of serialized context. Its truncation flags tell you what did not fit. Agents can select root nodes, traversal direction, relation types and property names through the API or query_graph_context MCP tool. Retrieval expands node IDs before loading properties, so large payloads do not inflate the traversal. Results have stable ordering for unchanged data and carry current revisions. The response includes serialized_bytes, budget, and truncated flags for nodes, relationships, properties and traversal expansion. A large omitted payload is marked properties_omitted: true. Choose an empty property list for metadata only. Execution traces are excluded unless explicitly requested.

Combine knowledge with workflows

Choose Workflow access, select a project, and grant Read graph context, Queue graph changes, or both. This is a grant to the entire workspace graph, including table records from other projects. It is separate from per-table workflow permissions. Project-scoped API keys cannot access the graph directly. In the workflow builder, add a Knowledge graph step:
  • Fetch agent context takes an array of node UUIDs or a templated array, plus depth, node and byte budgets. Place an AI or Agent step after it and use {{graph_result}} in that step’s input. Select only the properties it needs.
  • Queue knowledge changes takes an array of graph operations. The workflow supplies stable mutation IDs and generates missing IDs for new nodes, so a retried step does not create duplicates. Its result contains receipt IDs.
Example context step:
Example mutation step after an AI call:
Queued changes normally consolidate after about 15 seconds. A queue receipt means the change is durably queued, not yet applied. Outages and backlog can delay it. Poll the mutation receipt before depending on the changed data; the next workflow step may run before the queue flushes.

Review human and agent conflicts

Human edits save immediately against the revision displayed when the edit began. Agent and workflow mutations supply their expected revision. If someone else changed the node first, the proposed values are saved as a conflict and the current record stays intact. Open Conflicts → Review proposal to compare the latest record with the proposal. Edit the complete properties, confirm you reviewed that revision, and choose Apply reviewed change. Or select Discard this proposal to keep the current record. If another edit arrives while you review, fetch the latest revision and review again. A stale write never silently wins. Queue shows pending or errored mutations and lets writers cancel changes that have not begun applying. A workspace suspension or downgrade pauses queued writes. Revoking a project’s graph grant prevents its queued changes from applying.

Inspect execution lineage

Execution displays recent workflow runs as a React Flow canvas. Select a run to see task states, attempts, sequence and branch connections. It refreshes every five seconds from workflow history. Click a task to inspect its metadata. The graph projection is asynchronous. It records execution metadata without copying prompts, payloads or HTTP headers. Large runs show a bounded portion of their history with a truncation notice; use the workflow’s history for more detail.

Connected agents and the API

Connected agents use these MCP tools through their existing workspace read/write grants: Node discovery defaults to included data. Curation tools can explicitly request membership: "excluded" or "all" when listing nodes, or include_excluded: true when reading a node. Context queries always honor inclusion. To change a choice, send kind: "table" or "node", its resource_id, boolean included, and the current inclusion expected_version (zero if no choice has been saved). Table versions are in the overview; node versions are in node.membership.version. They are independent of the record’s content revision. The API lives under /admin/api/data-workspace/graph. GET /overview and POST /inclusion expose the same controls. Queued operations require stable mutation_id and node_id UUIDs. Updates replace the complete properties object and require expected_version. Supported operations are create_node, update_node, delete_node, upsert_edge, and delete_edge. For an edge, node_id is the source; provide target_id and relation_type. Revision zero creates an edge; a positive current revision updates or removes one.

Limits and data lifecycle

Each queue request supports 25 operations and 256 KiB. A workspace can hold 500 pending operations or 16 MiB of queued proposals. Each node’s properties support 64 KiB and 20 levels of JSON nesting. Nodes can have up to 256 incoming and 256 outgoing active relationships. Pro supports 50,000 standalone/trace nodes; Team supports 250,000, separately from table-record allowances. Database storage limits also apply. System traces and completed/resolved receipts become eligible for cleanup after 30 days. Curated knowledge and unresolved conflicts persist. Workspace export includes graph records, relationships, inclusion choices, receipts and pending proposals, including after downgrade. Workspace deletion clears both the dedicated database and its pending mutation queue. This V1 provides structured graph context. Semantic/vector search and historical context snapshots are future capabilities.