Knowledge

Knowledge is where your workspace stores and organizes information. It's the data layer that your AI agents draw from and write to during conversations — and where your team manages records and documents.

Everything here is accessible from Knowledge in the sidebar.

Forms

Structured data collection powered by AI conversation. Define fields, attach forms to workflows, and your agent asks for the information naturally.

• 14 field types: text, email, phone, numbers, dates, dropdowns, multi-select, formulas, and more
• Single-record mode: One record per person (profiles, preferences, intake forms)
• Collection mode: Multiple records per person (appointments, assessments, orders)
• Custom validation with CEL expressions
• Bulk CSV import and export up to 10,000 records

Example: A "Patient Screening" form with fields for demographics, medical history, and eligibility criteria — the AI collects it all through natural conversation.

Files

Document and file storage with folder organization, access controls, and AI-powered search.

• Drag-and-drop upload with nested folder hierarchies
• Access levels: private, users-only, or public
• Semantic indexing for AI-powered document search
• Agents can read, search, and create files during conversations

Example: A "FAQ Documents" folder that your support agent searches when answering customer questions.

Labels

Color-coded tags for organizing members, chats, and files.

• Three independent label systems: member labels, chat labels, and file labels
• Drive automations using CEL expressions ("vip" in member.labels)
• Filter and segment members, conversations, and documents
• Auto-assign labels through actions or chat summary generation

Example: Label members as "vip", "needs-followup", or "intake-complete" to drive routing, reporting, and automation.

Computed Fields

Auto-calculated formula fields for scoring, totals, and derived values.

• CEL expressions that compute values from other form fields
• Recalculate automatically when dependent fields change
• Common uses: clinical scoring (GAD-7, PHQ-9), BMI, order totals, age calculations

Member Memory

Persistent facts that agents remember about a member across conversations. Operators can view, add, edit, and delete memories from the Memory tab on a member's profile page.

What the Memory tab shows

Each memory has:

• Key — a short slug identifier (e.g. prefers-morning, allergic-penicillin)
• Description — a one-line summary always loaded into the agent's prompt
• Content — optional longer detail the agent can retrieve on demand

Memories are displayed in priority order (higher priority = loaded first). The creation and last-updated timestamps are shown for each entry.

Managing memories as an operator

• Add: click Add Memory, enter a key and description, and optionally provide extended content.
• Edit: click any memory row to update the description, content, or priority.
• Delete: use the delete action on a row. Deletion is permanent for the live record, but an immutable audit version is preserved in the version history for compliance purposes.

Internal memory vs AgentCore memory

The Memory tab shows two distinct memory systems:

	Internal Memory	AgentCore Memory
Storage	Workspace PostgreSQL	AWS Bedrock
Interface	Explicit tools (save/delete/read)	Automatic extraction
Visibility	Operator can view/edit via UI/API	Opaque cloud service
Feature gate	None	AGENT_MEMORY workspace feature
Versioning	Full version history kept	Not retained locally

Both types are surfaced together in the Memory tab so operators have a unified view of what the agent knows about a member.

PHI handling

Memory content is classified as Protected Health Information (PHI). It may contain health conditions, treatment preferences, or clinical notes. Gravity Rail:

• Stores internal memories encrypted at rest in your workspace database
• Audit-logs all create, update, and delete operations per HIPAA §164.312(b)
• Never logs the memory content itself — only the key slug appears in audit records
• Never exports AgentCore memory content to third-party systems beyond AWS Bedrock

Operators should apply the same care when entering memory content as they would for any other clinical record.