Memory is what your agent remembers between messages, between conversations, and across channels. Getting it right is the difference between an agent that feels alive and one that feels like a goldfish.
Three levels of memory
1. Within a single conversation
Always on. Inside one back-and-forth, the agent remembers everything the customer has said. This is the context window — controlled by the model, not by you.
2. Across conversations with the same customer
Configurable under Settings → Memory. When on, the agent recognises a returning customer (by phone, email, or session ID) and brings forward what it learned about them last time.
- Their preferences — language, tone, frequently asked topics
- Their history — past purchases, past tickets, previously stated facts
- Their identity — name, email, account number once collected
3. Long-term knowledge (sources)
Things that are true about your business, not about a specific customer. These live in Sources — see the Sources overview.
Memory settings
- Memory type — how memories are stored. Most teams use the default.
- Context window size — how many tokens of past conversation the agent considers. Larger = more context, slightly more expensive per message.
- State persistence — whether interrupted conversations resume where they left off.
How customer identity is matched
Memory only works if Chatzuri can recognise that two messages came from the same person. Identity is matched by, in order:
- Authenticated user ID (if you pass one in the embed)
- Phone number (WhatsApp, SMS, voice channels)
- Email address (Email channel, lead form)
- Session ID stored in the browser (Website widget)
A customer who messages from WhatsApp and then visits the website widget anonymously will be treated as two different people until you collect a shared identifier (an email, for example).