Changelog

xAI / Grok Now Supported

We’ve added xAI support in the latest SDK version. To enable xAI models, set your XAI_API_KEY as an environment variable: export XAI_API_KEY="...".

Core Memory and Archival Memory SDK APIs Renamed to Blocks and Passages

Given the confusion around our advanced functionality for managing memory, we’ve renamed the Core Memory SDK API to blocks and the Archival Memory SDK API to passages so that our API naming reflects the unit of memory stored. This change only affects our SDK, and does not affect Letta’s Rest API.

Before

1
from letta_client import CreateBlock, Letta
2
client = Letta(
3
  token="YOUR_API_KEY",
4
)
5
agent = client.agents.create(
6
  model="gpt-4o-mini",
7
  embedding="openai/text-embedding-ada-002"
8
  memory_blocks=[
9
    CreateBlock(
10
      "label": "human",
11
      "value": "name: Caren"
12
    ),
13
  ],
14
)
15
blocks = client.agents.core_memory.list_blocks(agent_id=agent.id)
16
client.agents.core_memory.detach_block(agent_id=agent.id, block_id=blocks[0].id)

After

1
from letta_client import CreateBlock, Letta
2
client = Letta(
3
  token="YOUR_API_KEY",
4
)
5
agent = client.agents.create(
6
  model="gpt-4o-mini",
7
  embedding="openai/text-embedding-ada-002"
8
  memory_blocks=[
9
    CreateBlock(
10
      "label": "human",
11
      "value": "name: Caren"
12
    ),
13
  ],
14
)
15
blocks = client.agents.blocks.list(agent_id=agent.id)
16
client.agents.blocks.detach(agent_id=agent.id, block_id=blocks[0].id)

New Identities Feature

We’ve added a new Identities feature that helps you manage users in your multi-user Letta application. Each Identity can represent a user or organization in your system and store their metadata.

You can associate an Identity with one or more agents, making it easy to track which agents belong to which users. Agents can also be associated with multiple identities, enabling shared access across different users. This release includes full CRUD (Create, Read, Update, Delete) operations for managing Identities through our API.

For more information on usage, visit our Identities documentation and usage guide.

Project Slug Moved to Request Header

Project slug can now be specified via request header X-Project for agent creation. The existing project parameter will soon be deprecated.

Before

1
curl -X POST https://app.letta.com/v1/agents \
2
  -H 'Content-Type: application/json' \
3
  -H 'Authorization: Bearer YOUR_API_KEY' \
4
  -d '{
5
    "project":"YOUR_PROJECT_SLUG"
6
    "model":"gpt-4o-mini",
7
    "embedding":"openai/text-embedding-ada-002"
8
    "memory_blocks": [
9
      {
10
        "label": "human",
11
        "value": "name: Caren"
12
      }
13
    ],
14
  }'

After

1
curl -X POST https://app.letta.com/v1/agents \
2
  -H 'Content-Type: application/json' \
3
  -H 'Authorization: Bearer YOUR_API_KEY' \
4
  -H 'X-Project: YOUR_PROJECT_SLUG' \
5
  -d '{
6
    "model":"gpt-4o-mini",
7
    "embedding":"openai/text-embedding-ada-002"
8
    "memory_blocks": [
9
      {
10
        "label": "human",
11
        "value": "name: Caren"
12
      }
13
    ],
14
  }'

New Features

Google Vertex support

Google Vertex is now a supported endpoint type for Letta agents.

Option to disable message persistence for a given agent

Letta agents now have an optional message_buffer_autoclear flag. If set to True (default False), the message history will not be persisted in-context between requests (though the agent will still have access to core, archival, and recall memory).

Temperature and Max Tokens Supported via LLM Config

These values are now configurable when creating and modifying agents via llm_config parameter for subsequent LLM requests.

Agents API Improvements

Agent Search

The /v1/agents/search API has been updated to support pagination via after query parameter

Agent Creation from Template

The /v1/templates/ creation API has been updated to support adding tags at creation time

Query tools by name

The List Tools API now supports querying by tool name.

1
send_message_tool_id = client.agents.tools.list(tool_name="secret_message")[0].id

Authorization header now supports password

For self-deployed instances of Letta that are password-protected, the Authorization header now supports parsing passwords in addition to API keys. X-BARE-PASSWORD will still be supported as legacy, but will be deprecated in a future release.

Before:

1
curl --request POST \
2
  --url https://MYSERVER.up.railway.app/v1/agents/ \
3
  --header 'X-BARE-PASSWORD: password banana' \
4
  --header 'Content-Type: application/json' \
5
  --data '{
6
    ...
7
  }'

After:

1
curl --request POST \
2
  --url https://MYSERVER.up.railway.app/v1/agents/ \
3
  --header 'AUTHORIZATION: Bearer banana' \
4
  --header 'Content-Type: application/json' \
5
  --data '{
6
    ...
7
  }'

Password can now be passed via the token field when initializing the Letta client:

1
client = LettaClient(
2
    base_url="https://MYSERVER.up.railway.app",
3
    token="banana",
4
)

Tool rules improvements

ToolRule objects no longer should specify a type at instantiation, as this field is now immutable.

Before:

1
  rule = InitToolRule(
2
    tool_name="secret_message",
3
    type="run_first"
4
)

After:

1
  rule = InitToolRule(tool_name="secret_message")

Letta also now supports smarter retry behavior for tool rules in the case of unrecoverable failures.

New API routes to query agent steps

The List Steps and Retrieve Step routes have been added to enable querying for additional metadata around agent execution.

Consistency Across Messages APIs

Flattened UserMessage content

The content field on UserMessage objects returned by our Messages endpoints have been simplified to flat strings containing raw message text, rather than JSON strings with message text nested inside.

Before:

1
  {
2
    "id": "message-dea2ceab-0863-44ea-86dc-70cf02c05946",
3
    "date": "2025-01-28T01:18:18+00:00",
4
    "message_type": "user_message",
5
    "content": "{\n  \"type\": \"user_message\",\n  \"message\": \"Hello, how are you?\",\n  \"time\": \"2025-01-28 01:18:18 AM UTC+0000\"\n}"
6
  }

After:

1
  {
2
    "id": "message-dea2ceab-0863-44ea-86dc-70cf02c05946",
3
    "date": "2025-01-28T01:18:18+00:00",
4
    "message_type": "user_message",
5
    "content": "Hello, how are you?"
6
  }

Top-level use_assistant_message parameter defaults to True

All message related APIs now include a top-level use_assistant_message parameter, which defaults to True if not specified. This parameter controls whether the endpoint should parse specific tool call arguments (default send_message) as AssistantMessage objects rather than ToolCallMessage objects.

Before:

1
response = client.agents.messages.create(
2
    agent_id=agent.id,
3
    messages=[
4
        MessageCreate(
5
            role="user",
6
            content="call the big_return function",
7
        ),
8
    ],
9
    config=LettaRequestConfig(use_assistant_message=False),
10
)

After:

1
response = client.agents.messages.create(
2
    agent_id=agent.id,
3
    messages=[
4
        MessageCreate(
5
            role="user",
6
            content="call the big_return function",
7
        ),
8
    ],
9
    use_assistant_message=False,
10
)

Previously, the List Messages endpoint defaulted to False internally, so this change may cause unexpected behavior in your code. To fix this, you can set the use_assistant_message parameter to False in your request.

1
messages = client.agents.messages.list(
2
    limit=10,
3
    use_assistant_message=False,
4
)

Consistent message return type

All message related APIs return LettaMessage objects now, which are simplified versions of Message objects stored in the database backend. Previously, our List Messages endpoint returned Message objects by default, which is no longer an option.