Agent memory

Agent Memory is an experimental feature that allows your assistant to persist data about users across conversations. This enables experiences like repeat-caller logic, customer preferences, and cross-channel continuity.

It uses a key-value store attached to a user identifier (like phone number). Memory is read at the start of each turn and written after the conversation ends.

You can:

• Read memory using conv.memory.get("key")
• Write memory by setting conv.state["key"] = value (if configured via state_keys)

​

Configuration

​

JSON config (Agent Studio)

To enable Agent Memory:

{
  "memory": {
    "repeat_caller": {
      "analytics_enabled": true,
      "state_keys": ["booking_day", "cheese_type"],
      "identifier_blacklist": ["+447777"]
    }
  }
}

• analytics_enabled: Adds repeat caller metrics to Studio analytics
• state_keys: Keys to save from conv.state to memory
• identifier_blacklist: Optional list of excluded IDs (e.g. test numbers)

​

Using memory in functions

​

Read memory

def start_function(conv: Conversation):
  if "cheese_type" in conv.memory:
    cheese = conv.memory["cheese_type"]
    conv.say(f"You previously picked {cheese}.")

​

Write memory (via conv.state)

def book(conv: Conversation):
  conv.state["cheese_type"] = "brie"

​

Example: conditional reuse

def start_function(conv: Conversation):
  booking_day = conv.memory.get("booking_day")
  cheese_type = conv.memory.get("cheese_type")
  if booking_day and cheese_type:
    conv.say(f"You booked {cheese_type} for {booking_day}. Rebook?")

​

Platform (context.memory)

If you’re working in the agent worker (not Studio), use context.memory:

context.memory.patch_builder_memory(
  identifier=context.turn.user_input.from_,
  fields={"cheese_type": "gouda"}
)

You can also:

• get_builder_memory(...)
• link_identifiers(...)

​

Memory behaviour

• Memory is read once per turn and cached.
• Memory is persisted at the end of a conversation.
• Only supported identifier today: phone number.

​

FAQ

​

Can I store nested JSON?

No. Each value must be a string, number, or flat serializable object. Always avoid packing PII into a single field.

​

Can I check if memory exists?

Yes:

if "booking_day" in conv.memory:

​

Can I prevent overwrites?

Not directly, but you can check before writing:

if "cheese_type" not in conv.memory:
  conv.state["cheese_type"] = value

​

Can I link identifiers?

Yes. This is supported by context.memory.link_identifiers().

Need help? Contact platform-support@poly-ai.com.