Letta (MemGPT) Integration

THSP safety validation for Letta agents with persistent memory.

Letta (formerly MemGPT) is a platform for building stateful AI agents with self-editing memory. This integration adds safety validation at message input, tool execution, and memory operations.

Installation

pip install letta-client sentinelseed

Quick Start

Method 1: Wrapped Client

from letta_client import Letta
from sentinelseed.integrations.letta import SentinelLettaClient

base = Letta(api_key="your-letta-key")

client = SentinelLettaClient(
    base,
    api_key="your-openai-key",
    mode="block",  # block, flag, log
)

agent = client.agents.create(
    model="openai/gpt-4o-mini",
    memory_blocks=[
        {"label": "human", "value": "User info"},
        {"label": "persona", "value": "AI assistant"},
    ],
)

response = client.agents.messages(agent.id).create(
    input="Hello, how are you?"
)

Method 2: Safety Tool

from letta_client import Letta
from sentinelseed.integrations.letta import create_sentinel_tool

client = Letta(api_key="your-key")

tool = create_sentinel_tool(
    client,
    api_key="your-openai-key",
    require_approval=True,
)

agent = client.agents.create(
    model="openai/gpt-4o-mini",
    tools=[tool.name],
    memory_blocks=[...],
)

Method 3: Safe Agent Factory

from sentinelseed.integrations.letta import create_safe_agent

agent = create_safe_agent(
    client,
    validator_api_key="your-openai-key",
    model="openai/gpt-4o-mini",
    memory_blocks=[...],
    include_safety_tool=True,
    high_risk_tools=["web_search", "run_code"],
)

Features

Message Validation

• Input validation: Check user messages before processing
• Output validation: Check agent responses before returning
• Configurable modes: block, flag, or log

Memory Integrity

from sentinelseed.integrations.letta import create_memory_guard_tool

guard = create_memory_guard_tool(
    client,
    secret="your-hmac-secret",
)

Memory Content Validation (v2.0)

Validates content BEFORE HMAC signing, detecting injection attacks at the source.

Detected Patterns:

Approval Handler

from sentinelseed.integrations.letta import sentinel_approval_handler

decision = sentinel_approval_handler(
    approval_request={
        "tool_name": "run_code",
        "arguments": {"code": "print('hello')"},
    },
    api_key="your-openai-key",
    auto_approve_safe=True,
    auto_deny_unsafe=True,
)

THSP Gates

Links

• Letta Documentation