THE FIELD GUIDE · 3 min read
Knowledge, memory, and human control overview
Understand how Brain knowledge, agent memory, escalations, and workflow approvals relate.
Venture Brain
Brain is the venture knowledge surface. It stores structured knowledge and memory produced or used by agents. Users can browse entries, inspect details, search and filter the knowledge set, and review how the venture's operating context evolves.
Knowledge may include facts, decisions, lessons, research, and workflow-derived context. Correct source information when it changes; stale memory can influence later agent work.
Brain is scoped by organization and venture. Workflows mount or retrieve only the knowledge and resource contexts authorized for that run. Retrieval is context, not a command: the workflow should still treat the canonical repository, provider, strategy record, or project-management item as authoritative.
For observations, versioning, schedules, thresholds, freezing, models, analytics, and synthesis cost, see Brain and synthesis.
Memory
Agents can use persistent memory primitives to retain useful information across work. Memory is scoped and authorization-checked. It is not a replacement for the canonical strategy, repository, or project-management system.
Review important memory against its canonical source. Correct the source first, then record the correction so later synthesis can incorporate it.
Escalations
CXO agents create durable escalation records when an off-platform blocker needs CEO or human help. Pending actions expose these records in the product.
An escalation may ask for:
- clarification or missing context;
- approval of a consequential action;
- a selection between alternatives;
- review of a workflow output;
- operational intervention after repeated failure.
Resolve escalation records from the venture Inbox or Needs you queue. The response becomes a high-importance observation for later agent use.
An escalation record is not automatically a signal to a waiting workflow. Workflow-specific approval and input controls continue their associated run; generic escalation resolution does not guarantee that behavior.
For categories, urgency, ownership, structured responses, retirement, and 14-day auto-expiry, see Escalations and human control.
Approval controls
The venture Inbox brings pending decisions into one ranked list, including workflow input, escalations, content, and email approvals. Each item retains the controls for the decision it owns. You can also use the originating run for its available actions. Finished development work uses its saved task title, falling back to the pull request’s title when needed, and its age shows how long the current decision has been waiting.
Use Chat about this (or Ask about this in compact views) where offered to open a new conversation with the relevant teammate and the item’s context. Decision history shows recorded outcomes, who decided, when, and any reason, with a link to the original item when it is still available.
The Inbox keeps your filter, selection, and list position when navigating away and back. On mobile, open an item to move from the list to its detail, then return to the list.
On Home, Needs you shows the same pending decisions as the Inbox, in the same order, one at a time. Use Previous, Next, or Skip for now to browse; skipping does not resolve the item. Open the Inbox to browse the full list.
Finished work and expiring decisions
While you were away summarizes finished work since your previous visit above the Inbox decisions. Open a result from the summary or dismiss the summary; it is information, not another approval, and does not add to the pending-decision count. Chat-started results also return to the originating conversation.
Time-bound decisions show an expiry when one is set. Once their deadline passes, they leave the pending list and cannot be approved. Expiry does not authorize the action.
In the Emails filter, Approve all lets you confirm the pending emails together. Review the confirmation before sending, then check the per-email results: a failure or unconfirmed approval needs attention.
Agent actions use contextual permissions and guardrails. See Agent tool access (MCP).
Agent tool access (MCP)
Understand how Organ gives agents a task-specific, permissioned toolset and how to diagnose unavailable actions.
Getting credentials and tool access as an agent
Get credentials inside a dispatched task via resourceContextIds and the grant_resources scope, or during a chat session via request_resource_access, get_credential_status, and list_resource_contexts.