letta-conversations-api（letta conversations API 文档）

Letta Conversations API

The Conversations API allows multiple isolated message threads on a single agent. Each conversation maintains its own message history while sharing the agent's memory blocks and tools.

When to Use This Skill

• Building multi-user chat applications (each user gets their own conversation)
• Implementing session management with separate contexts
• A/B testing agent responses across isolated conversations
• Any scenario where you need multiple independent chat threads with one agent

Key Concepts

Python SDK Usage

Setup

from letta_client import Letta

client = Letta(base_url="https://api.letta.com", api_key="your-key")

Create a Conversation

conversation = client.conversations.create(agent_id="agent-xxx")
# conversation.id -> "conv-xxx"

Send Messages (Streaming)

stream = client.conversations.messages.create(
    conversation_id=conversation.id,
    messages=[{"role": "user", "content": "Hello!"}],
)

for msg in stream:
    if hasattr(msg, "message_type") and msg.message_type == "assistant_message":
        print(msg.content)

List Messages in a Conversation

messages = client.conversations.messages.list(
    conversation_id=conversation.id,
    limit=50,  # Optional: default 100
    after="message-xxx",  # Optional: cursor for pagination
    before="message-yyy",  # Optional: cursor for pagination
)

List All Conversations for an Agent

conversations = client.conversations.list(
    agent_id="agent-xxx",
    limit=50,  # Optional
    after="conv-xxx",  # Optional: cursor for pagination
)

Retrieve a Specific Conversation

conv = client.conversations.retrieve(conversation_id="conv-xxx")
# conv.in_context_message_ids -> list of message IDs in context window

REST API Endpoints

REST Example: Create and Send Message

# Create conversation
curl -X POST "https://api.letta.com/v1/conversations?agent_id=agent-xxx" \
  -H "Authorization: Bearer $LETTA_API_KEY" \
  -H "Content-Type: application/json"

# Send message (streaming response)
curl -X POST "https://api.letta.com/v1/conversations/conv-xxx/messages" \
  -H "Authorization: Bearer $LETTA_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Accept: text/event-stream" \
  -d '{"messages": [{"role": "user", "content": "Hello!"}]}'

Conversation Schema

class Conversation:
    id: str                      # "conv-xxx"
    agent_id: str                # Associated agent ID
    created_at: datetime         # Creation timestamp
    summary: Optional[str]       # Optional conversation summary
    in_context_message_ids: List[str]  # Message IDs in context window

Common Patterns

Multi-User Chat Application

# Each user gets their own conversation
user_conversations = {}

def get_or_create_conversation(user_id: str, agent_id: str) -> str:
    if user_id not in user_conversations:
        conv = client.conversations.create(agent_id=agent_id)
        user_conversations[user_id] = conv.id
    return user_conversations[user_id]

def send_user_message(user_id: str, agent_id: str, message: str):
    conv_id = get_or_create_conversation(user_id, agent_id)
    return client.conversations.messages.create(
        conversation_id=conv_id,
        messages=[{"role": "user", "content": message}],
    )

Paginating Through Message History

def get_all_messages(conversation_id: str):
    all_messages = []
    after = None

    while True:
        batch = client.conversations.messages.list(
            conversation_id=conversation_id,
            limit=100,
            after=after,
        )
        if not batch:
            break
        all_messages.extend(batch)
        after = batch[-1].id

    return all_messages

Important Notes

1. Streaming by default: The messages.create endpoint always streams responses
2. Shared memory: Memory block updates in one conversation are visible in all conversations for that agent
3. Message isolation: Conversation message history is completely isolated between conversations
4. Pagination: Use after/before cursors for efficient pagination, not offsets

Example Scripts

This skill includes two example scripts in the scripts/ directory:

1. conversations_demo.py - Comprehensive demo showing all API features

- Basic conversation flow - Conversation isolation testing - Listing and retrieving conversations - Pagination examples - Shared memory demonstration

1. conversations_cli.py - Interactive TUI for managing conversations

- Create/switch between conversations - Send messages with streaming responses - View message history - Switch between agents

Running the Examples

# Run the demo script
LETTA_API_KEY=your-key uv run letta/conversations/scripts/conversations_demo.py

# Run the interactive CLI
LETTA_API_KEY=your-key uv run letta/conversations/scripts/conversations_cli.py

# CLI with specific agent
LETTA_API_KEY=your-key uv run letta/conversations/scripts/conversations_cli.py --agent agent-xxx

SDK Gotchas

• Paginated responses use .items to access the list: client.agents.list().items
• Auth parameter is api_key, not token: Letta(base_url=..., api_key=...)
• Message streams must be consumed (iterate or list()) to complete the request