Accounts and account memory

Accounts is the durable memory of GTM Brain. Instead of treating every campaign as a new spreadsheet, the system keeps one tenant-scoped history for each account domain.

Account list

Each row summarizes the latest known state:

Search matches account names and domains. State filters help isolate cold, contacted, replied, meeting, or suppressed accounts. The record grid uses bounded cursor pages and loads another page as the user approaches the end. Sort and filter state are bound into the cursor so a cursor from one query cannot be reused for another.

The grid is intentionally spreadsheet-like: the Account column remains pinned while the remaining tenant-selected columns scroll horizontally, headers remain visible during long reviews, and compact semantic tags distinguish industries, select values, and lifecycle state without relying on color alone. Company marks use bundled open-source brand marks when available, otherwise try the company's own public favicon before falling back to an accessible monogram. GTM Brain does not disclose prospect domains to a third-party logo service.

Imported CRM identities remain internal. When Lightfield supplies a company name but no public website, the list and account workspace show the company name plus Imported company record or Website not available in Lightfield. A provider key such as lightfield-acc_… is retained for reconciliation and audit, but is never presented as the company name or treated as a clickable domain.

Display controls show or hide active system and tenant-defined fields. Tenant administrators create, version, and retire those fields from Data model. Account operators edit the values inside the account workspace; the update is optimistic-concurrency protected and emits a governed event rather than mutating the list projection.

Open an account record

Select a row to open the shared responsive account workspace. Its views are:

• Overview — proactive suggestions, recipient, qualification, evidence, campaign artifacts, delivery touches, and opportunity state.
• Relationships — ownership, linked records, duplicate review, and reversible merge state.
• Work — account-bound tasks and team/private notes with a compact composer.
• Activity — a readable per-company timeline of imports, governed edits, signals, outreach, replies, and meetings. Expand Audit details only when the event, correlation, or action-invocation identity is needed for an investigation.

The timeline is projected from tenant-scoped Fabric Platform AssetEvents. Imported Lightfield records receive an Imported from Lightfield provenance entry even when they have not yet accumulated GTM signals or touches. Event payloads and secret material are not exposed in this projection.

Evidence is organized as canonical facts with expandable source records. Repeated provider collections of the same source revision appear once with first-seen, last-seen, and consolidated-count metadata. The underlying immutable evidence and events remain available to replay and audit tooling.

Read the decision context in this order:

1. Signals — what GTM Brain observed and when.
2. Verdicts — score, recommended play, and why now reasoning.
3. Drafts and touches — what was proposed, approved, and delivered.
4. Replies and outcomes — what happened after contact.
5. State changes — the governed result of those actions.

This sequence separates evidence from model judgment. A high score is not itself a signal; it is a verdict derived from signals, ICP configuration, and account history.

Creating a task or note and changing an opportunity stage never writes the account projection directly. The UI submits a compact idempotent command to Temporal; Fabric Platform evaluates policy and state, emits the canonical AssetEvent, and the account workspace refreshes from the resulting projection.

Agent API

The protected product API exposes cursor-paged records at GET /gtm/accounts/records and the same focused projection at GET /gtm/accounts/history. Registered external agents use the curated resource returned by the external catalog:

GET /api/agents/gtm/{organizationId}/accounts/{account}/context
    ?registrationId={registrationId}

That resource requires the tenant agent key, an enabled registration, and the gtm.read_account_context grant. Agents propose account notes and tasks with POST /api/agents/gtm/{organizationId}/invoke; both mutations traverse the same Temporal and Fabric Platform runtime as the UI. The catalog does not advertise send, approval, publication, or campaign-activation actions.

Example: decide whether to approve a second touch

Suppose an account has:

Signal: Series B announced 12 days ago
Prior touch: first_touch sent 8 days ago
Outcome: no_reply
New signal: hiring Director of Data Platform today
Judge: score 88, play follow_up

Before approving:

• confirm the new job signal is distinct from the earlier funding trigger;
• verify the contact and domain are not suppressed;
• read the prior message so the follow-up does not repeat it;
• ensure the proposed draft explicitly uses the new hiring trigger.

Account states are governed

Operators do not drag accounts between columns. State changes are produced by actions and outcomes:

This prevents the UI from claiming an outcome that has no corresponding event or audit record.

Multi-tenant behavior

The same domain can exist in multiple organizations with disjoint histories. Every account read verifies organization membership and uses an organization-scoped store; users cannot query another tenant by changing the domain or workflow ID.

Next: Signals, Pipeline, and Contacts and opportunities.