Conversation Memory
Great conversations don't start from zero.
Conversation Memory helps your agents remember previous interactions, so every new conversation feels like a continuation instead of a fresh start.
Whether a Guest calls, sends a WhatsApp message, or replies by SMS, your agent remembers the context and responds accordingly.
Summary
On this page you'll learn how to:
- Enable memory for an agent.
- Choose which communication channels contribute to memory.
- Configure recent history and long-term memory.
- Review or clear a Guest's stored memory.
- Understand how Conversation Memory works behind the scenes — see Technical details.
What you can do
Conversation Memory lets your agents build context over time.
You can:
- Remember previous conversations across channels.
- Personalize future conversations automatically.
- Keep track of preferences, commitments, and ongoing requests.
- Continue conversations without asking Guests to repeat themselves.
- Review or clear stored memory for any Guest.
Why it matters for your business
People expect to be remembered.
Repeating the same information every time is frustrating — for Guests and for your team.
Conversation Memory helps your business:
- Save money: fewer repeated questions and shorter follow-ups.
- Work faster: agents start with context instead of starting over.
- Standardize service: prior interactions inform the next conversation consistently across channels.
- Delight every Guest: conversations feel continuous instead of reset.
Where to find it
Open:
Agent → Memory (/assistant/memory)
Top navigation order:
Inbound → Connections → Knowledge → Agent settings → Memory → Outbound
This page controls memory for the entire agent.
You no longer need to configure memory separately for WhatsApp or SMS — those channel sheets link here instead.
How memory works
Conversation Memory combines two types of context.
flowchart LR Sources[Phone WhatsApp SMS] Recent[Recent history] Summary[Long-term summary] Agent[Agent reply or call] Sources --> Recent Sources --> Summary Recent --> Agent Summary --> Agent
Recent history
Recent conversations within your selected look-back window.
This helps the agent understand what happened recently.
Long-term summary
A compact summary that preserves important information long after individual conversations fall outside the look-back window.
Examples include:
- preferences
- previous questions
- reservations
- commitments
- ongoing issues
- upcoming appointments or stays
Together, these two layers allow the agent to continue conversations naturally without loading every previous interaction.
Memory settings
Use memory
Turns Conversation Memory on or off for this agent.
When disabled, previous Guest context is not included in conversations.
Sources
Choose which channels contribute to memory.
Supported sources:
- Phone calls
- WhatsApp messages
- WhatsApp calls
- SMS
By default, all sources are enabled.
Recent history
Choose how much recent context to include.
Settings:
- Look back: 24 hours, 7 days, 30 days, or All time
- Recent messages: number of recent conversation turns to include (default: 20)
Long-term summary
When enabled, Voicetta updates a compact Guest summary after conversations.
This helps preserve important information beyond the recent history window.
This step needs your LLM API key (OpenAI by default). See Requirements.
Advanced settings let you customize:
- summary model
- temperature
- maximum tokens
Most workspaces can keep the default configuration.
Review a Guest
Search by phone number to:
- view the current memory summary
- clear stored memory
Clearing memory removes the summary and scheduled facts for that Guest while keeping the complete Conversation History.
Defaults
| Setting | Default |
|---|---|
| Use memory | On |
| Sources | Phone, WhatsApp messages, WhatsApp calls, SMS |
| Look back | 7 days |
| Recent messages | 20 |
| Long-term summary | On |
Phone-only agents use the same Memory page — you do not need SMS or WhatsApp connected to configure memory.
How it works across channels
Conversation Memory follows the Guest, not the channel.
| Channel | Memory behavior |
|---|---|
| Phone | Context is loaded before the call begins. |
| WhatsApp messages | Previous conversations help generate the next reply. |
| WhatsApp calls | Works the same way as phone calls. |
| SMS | Uses the same conversation memory as other messaging channels. |
One Memory configuration applies to every supported channel for that agent.
flowchart TB Settings[Agent Memory settings] Phone[Phone calls] WaText[WhatsApp messages] WaCall[WhatsApp calls] Sms[SMS] Settings --> Phone Settings --> WaText Settings --> WaCall Settings --> Sms
Requirements
Long-term summaries require your own LLM API key (Bring Your Own Key).
By default, Voicetta uses OpenAI GPT-4o mini to generate and refresh memory summaries.
Configure your API key in:
Settings → AI & API keys
See API Keys and BYOK.
If you select a different provider under Advanced, you'll need an API key for that provider instead.
Without a compatible API key:
- Conversation Memory can still use recent history.
- Long-term summaries won't be updated.
Technical details
For teams that want to understand how Conversation Memory works under the hood.
Scope
Conversation Memory is configured per agent.
Long-term summaries are stored per Guest phone number and isolated between agents.
Two agents in the same workspace never share Guest memory.
flowchart TB Guest[Guest phone number] AgentA[Agent A summary] AgentB[Agent B summary] Guest --> AgentA Guest --> AgentB
Memory lifecycle
After an eligible conversation finishes:
- Voicetta loads the existing Guest summary for this agent.
- The latest conversation is evaluated.
- The summary is updated when needed (requires a summary LLM key).
- The updated summary becomes available for future conversations.
Only the compact summary is carried forward — not entire conversation transcripts.
flowchart LR End[Conversation ends] Load[Load existing summary] Eval[Evaluate latest turns] Update[Update summary and facts] Next[Available on next channel] End --> Load Load --> Eval Eval --> Update Update --> Next
Load path at conversation time
When a new message or call starts:
- Resolve this agent's Memory settings.
- Load recent history from allowed sources within the look-back window.
- Load the lasting summary for this Guest + agent (if enabled).
- Inject that context into the text reply or voice agent prompt.
flowchart TD Start[New call or message] Resolve[Resolve agent Memory settings] Hist[Load recent history] Sum[Load lasting summary] Inject[Inject into reply or call prompt] Start --> Resolve Resolve --> Hist Resolve --> Sum Hist --> Inject Sum --> Inject
What is remembered
Conversation Memory may include:
- customer preferences
- previous questions
- reservation details
- ongoing issues
- commitments
- scheduled events
- important business context
It is designed to preserve operational context rather than act as a CRM.
Safety
Conversation Memory should not replace identity verification.
Agents should always confirm information such as:
- payment details
- personal information
- booking dates
- any action where mistakes could impact the Guest or your business
Privacy
Conversation Memory is intended to improve future conversations.
It stores operational context for each Guest while keeping Conversation History separate.
Deleting memory removes the stored summary only. Conversation History remains available unless deleted separately.