Skip to main content

​
The problem

Without context, each agent run is stateless. When a guest asks “I need towels for room B12” and the concierge replies “Large or small?”, the next message “medium” means nothing to the agent — it has no memory of the previous exchange.

​
How it works

WhatsRB automatically injects conversation context into each agent run, so the agent understands the conversation history:
  1. After each run, the agent generates a context_summary — a one-line snapshot of the conversation state (intent, extracted fields, pending questions).
  2. On the next inbound message from the same phone number, WhatsRB injects the previous run’s context_summary and last_agent_output (intent + actions) into the agent’s input.
  3. The agent uses this context to correctly interpret follow-up messages.
Context is scoped per agent + phone number. Two different agents or two different phone numbers never share context.

​
Configuration

Control context duration with context_ttl_minutes on the agent:

​
What gets injected

When a previous run exists within the TTL window, the agent receives two extra fields in its input:

​
Example flow

Without context, message 2 would be classified as unclear or misinterpreted. With context, the agent knows the guest is specifying towel size for room B12.

​
Interaction with Human Takeover

When a human takeover pause is active, the agent doesn’t run at all — so no context is generated during the pause. When the pause expires and the agent resumes, it picks up context from the last run before the pause (if still within TTL).

​
Best practices

  • Keep the default TTL (60 min) for most use cases. It covers natural conversation gaps without carrying stale context.
  • Increase to 120-240 min for async workflows where customers may reply hours later (e.g., B2B support).
  • Set to 0 for agents that handle fully independent requests (e.g., a pure classification agent with no follow-up expected).
  • The agent automatically generates context_summary — no prompt engineering needed. It’s built into the default tool schema.