Pre-Call Context

Pre-Call Context reuses what the platform already learned on a caller’s earlier calls. When it is on and the caller matches an earlier finished call of the same agent, the platform seeds previous_call_* variables before the prompt renders, so the greeting and the prompt can pick up where things left off. It is off by default, and the default configuration adds no latency to the call.

How it works

When a call ends, post-call analysis writes a continuation brief onto the run: the call summary, the flow node it left off on, and the open items (next steps) it found.
When the next call connects, one indexed read finds that caller’s recent finished calls for this agent, within your organization and within the lookback window. Purged calls never resurface.
The variables below are seeded from the most recent call (and every carried call is available as {{previous_calls}}). A value you supply for the current call always wins.
If nothing matches, or the lookup fails, is_returning_caller is false and the call starts fresh.

Configuration

Open the agent’s Pre-Call Context panel (a card in the conversational editor, a modal in the single-prompt editor), or set the pre_call_context block on workflow_configurations through the agent API.

Control	Stored as	What it does	Latency
Enable pre-call context	`enabled`	Off by default; an absent block reads as off.	none
Context source	`source`	`platform_history` (default) reads your own earlier calls. `webhook` asks your endpoint before the call. `both` merges the webhook result over platform history; the webhook wins on a shared key.	none / adds
Match returning callers by	`match_by`	`caller_number` (default) matches on the phone number. `contact_id` matches the caller’s saved contact, so a number spelt differently still matches, and falls back to the number.	none
Assembly mode	`mode`	`context_only` (default) seeds the stored variables. `llm_brief` spends one bounded LLM turn writing a short paragraph as `previous_call_brief`, and falls back to context only on any failure or timeout.	none / adds
Look back (days)	`lookback_days`	How far back an earlier call still counts. Default 30, range 1 to 365.	none
Prior calls to carry	`max_prior_calls`	How many recent calls are carried. Default 3, range 1 to 10. The most recent fills `previous_call_*`; every carried call is in `previous_calls`.	none
Resume conversation state	`resume_conversation_state`	Advanced. Restarts the agent at the flow node the earlier call left off on (voice and chat). Only the node is resumed, not the earlier transcript; if that node no longer exists the call starts at the beginning.	none
Webhook URL and credential	`webhook_url`, `webhook_credential_uuid`	Shown for `webhook` and `both`. The URL is required for those sources and must be public. See Pre-Call Data Fetch for the request and response format.	adds

{
  "pre_call_context": {
    "enabled": true,
    "source": "platform_history",
    "match_by": "contact_id",
    "mode": "context_only",
    "lookback_days": 30,
    "max_prior_calls": 3,
    "resume_conversation_state": false
  }
}

Variables the prompt can use

Variable	What it holds
`{{ is_returning_caller }}`	`true` when a matching earlier call was found, else `false`
`{{ previous_call_summary }}`	Summary of the most recent earlier call
`{{ previous_call_status }}`	How that call ended
`{{ previous_disconnection_reason }}`	Why it ended, if recorded
`{{ previous_call_at }}`	When it happened (ISO 8601)
`{{ previous_call_count }}`	How many earlier calls were carried (up to `max_prior_calls`)
`{{ previous_call_node }}`	Flow node it left off on, if tracked
`{{ previous_call_next_steps }}`	The open items it left for a follow-up, as a list of short phrases
`{{ previous_call.<field> }}`	Any field that call extracted, for example `{{ previous_call.appointment_date }}`
`{{ previous_calls }}`	Every carried call, most recent first, each with `status`, `at`, `summary`, `disconnection_reason`, `node` and `next_steps`
`{{ previous_call_brief }}`	The LLM-written paragraph, only in `llm_brief` mode

Testing from the editor

A test call from the editor has no telephone leg to identify the caller, so add a phone_number Test Input (the number you want to be recognised as). The editor, the embed widget and the text-chat tester all match on it, and stamp the contact so contact_id matching finds those runs too. On a real call the dialled or calling number is used.

Privacy and limits

Every lookup is scoped to your organization and to the same agent. A purged call is never reused. The LLM brief does not surface personal details unless they are essential to continue. Matching on a phone number cannot tell two people sharing a line apart; use contact_id matching where you maintain contacts, and keep prompts from reading earlier details out unprompted.

Getting started

Core Concepts

Configurations

Voice Agent Builder

Telephony

Integrations

Voice Studio

How it works

Configuration

Variables the prompt can use

Testing from the editor

Privacy and limits

​How it works

​Configuration

​Variables the prompt can use

​Testing from the editor

​Privacy and limits

How it works

Configuration

Variables the prompt can use

Testing from the editor

Privacy and limits