letta-app-development（莱塔应用程序开发）

Letta App Development Skill

Build applications on top of the Letta API — a model-agnostic, stateful API for building persistent agents with memory and long-term learning. The Letta API powers Letta Code and the Learning SDK. This skill covers the core patterns for creating agents, managing memory, building custom tools, and handling multi-user scenarios.

When to Use This Skill

• Building applications that need persistent, stateful AI agents
• Creating chatbots, assistants, or autonomous agents with memory
• Integrating Letta into existing web/mobile applications
• Building multi-user applications where each user has their own agent
• Understanding the API layer that Letta Code and Learning SDK are built on

Quick Start

See getting-started.md for first-time setup and common onboarding issues.

SDK Versions Tested

Examples last tested with:

• Python SDK: letta-client==1.7.1
• TypeScript SDK: @letta-ai/letta-client@1.7.1

Core Concepts

1. Client Setup

See client-setup.md for initialization patterns:

• Letta Cloud vs self-hosted connections
• Environment variable management
• Singleton patterns for web frameworks

2. Memory Architecture

See memory-architecture.md for memory patterns:

• Core Memory Blocks: Always in-context (persona, human, custom blocks)
• Archival Memory: Large corpus with semantic search
• Conversation History: Searchable message history
• Shared Blocks: Multi-agent coordination

3. Custom Tools

See custom-tools.md for tool creation:

• Simple function tools with auto-generated schemas
• Tools with environment variable secrets
• BaseTool class for complex schemas
• Sandboxed execution requirements

4. Client-Side Tools

See client-side-tools.md for local tool execution:

• Execute tools on your machine while agent runs on Letta API
• How Letta Code runs Bash/Read/Write locally
• Approval-based flow with type: "tool" responses
• Access local files, databases, and private APIs

5. Client Injection & Secrets

See client-injection.md for server-side tool patterns:

• Pre-injected client variable on Letta Cloud
• Building custom memory tools that modify agent state
• Agent secrets via os.getenv()
• LETTA_AGENT_ID for self-referential tools

6. Multi-User Patterns

See multi-user.md for scaling:

• One agent per user (personalization)
• Shared agent with Conversations API
• Identity system for user context

7. Streaming

See streaming.md for real-time responses:

• Basic SSE streaming
• Long-running operations with include_pings
• Background execution and resumable streams

8. Conversations

Conversations enable parallel sessions with shared memory:

• Thread-safe concurrent messaging (agents.messages.create is NOT thread-safe)
• Shared memory blocks across all conversations
• Separate context windows per conversation
• Use for: same user with multiple parallel tasks, multi-threaded applications

9. Sleeptime Agents

See sleeptime.md for background memory processing:

• Enable with enable_sleeptime=True
• Background agent refines memory between conversations
• Good for agents that learn over time

10. Agent Files & Folders

See agent-files.md for portability and file access:

• Export/import agents with .af files
• Attach folders to give agents document access
• Migration checklist for moving agents

11. Tool Rules

See tool-rules.md for constraining tool execution:

• InitToolRule - Force a tool to run first
• ChildToolRule - Control which tools can follow
• TerminalToolRule - End agent turn after tool
• Sequential pipelines and approval workflows

Quick Reference

Python SDK

pip install letta-client

from letta_client import Letta

# Cloud
client = Letta(api_key="LETTA_API_KEY")

# Self-hosted
client = Letta(base_url="http://localhost:8283")

TypeScript SDK

npm install @letta-ai/letta-client

import { Letta } from "@letta-ai/letta-client";

// Cloud
const client = new Letta({ apiKey: process.env.LETTA_API_KEY });

// Self-hosted
const client = new Letta({ baseUrl: "http://localhost:8283" });

Examples

See the examples/ directory for runnable code:

Python:

• 01_basic_client.py - Client initialization
• 02_create_agent.py - Agent creation with memory blocks
• 03_custom_tool_simple.py - Basic custom tool
• 04_custom_tool_secrets.py - Tool with environment variables
• 05_send_message.py - Basic messaging
• 06_send_message_stream.py - Streaming responses
• 07_multi_user.py - Multi-user patterns
• 08_archival_memory.py - Archival memory operations
• 09_shared_blocks.py - Multi-agent shared memory
• 10_conversations.py - Parallel sessions with conversations
• 11_client_injection.py - Custom memory tools with injected client
• 12_tool_rules.py - Constraining tool execution order
• 13_client_side_tools.py - Execute tools locally (like Letta Code)

TypeScript:

• 01_basic_client.ts - Client initialization
• 02_create_agent.ts - Agent creation
• 03_send_message.ts - Basic messaging
• 04_send_message_stream.ts - Streaming
• 05_nextjs_singleton.ts - Next.js pattern
• 06_multi_user.ts - Multi-user patterns
• 07_conversations.ts - Parallel sessions
• 08_custom_tool.ts - Custom tools with secrets
• 09_archival_memory.ts - Long-term storage
• 10_shared_blocks.ts - Multi-agent shared memory
• 11_client_injection.ts - Custom memory tools
• 12_tool_rules.ts - Tool execution order
• 13_client_side_tools.ts - Execute tools locally (like Letta Code)

Troubleshooting

Key Gotchas

1. Imports in tools must be inside the function - Tools run in a sandbox without access to top-level imports
2. Use os.getenv() for secrets - Don't pass sensitive data as function arguments
3. On Cloud, use injected client - Don't instantiate Letta() inside tools, use the pre-injected client
4. Memory blocks are character-limited - Use archival memory for large data
5. Streaming requires include_pings=True for long operations - Prevents timeout on Cloud
6. SDK 1.0 uses .update() not .modify() - Method was renamed
7. LETTA_AGENT_ID is always available - Use it in tools to reference the current agent
8. Archival tools need include_base_tools=True - Not attached by default
9. Use memory_insert for shared blocks - Safest for concurrent writes (append-only)
10. Tool docstrings require Args section - Parameters need descriptions or schema generation fails

TypeScript SDK Notes

// Client initialization uses baseURL (not baseUrl)
const client = new Letta({ apiKey: "...", baseURL: "http://localhost:8283" });

// Block API: positional args changed
client.agents.blocks.attach(blockId, { agent_id });      // blockId is first
client.agents.blocks.retrieve(blockLabel, { agent_id }); // label is first

// Passages.create returns array
const passages = await client.agents.passages.create(agentId, { text: "..." });
const passage = passages[0];

// Content can be string | array - use type guard
const content = typeof msg.content === "string" ? msg.content : JSON.stringify(msg.content);

// Conversations API returns streams by default
const stream = await client.conversations.messages.create(convId, { messages: [...] });
for await (const chunk of stream) { ... }

// Tool rule types
{ type: "run_first", tool_name: "..." }           // InitToolRule
{ type: "constrain_child_tools", tool_name: "...", children: [...] } // ChildToolRule
{ type: "exit_loop", tool_name: "..." }           // TerminalToolRule

Quick Reference

# Client
client = Letta(api_key=os.getenv("LETTA_API_KEY"))

# Create agent
agent = client.agents.create(
    model="anthropic/claude-sonnet-4-5-20250929",
    embedding="openai/text-embedding-3-small",
    memory_blocks=[{"label": "persona", "value": "..."}],
    include_base_tools=True,  # archival memory tools
    enable_sleeptime=True,    # background memory processing
)

# Send message
response = client.agents.messages.create(
    agent_id=agent.id,
    messages=[{"role": "user", "content": "Hello"}]
)

# Stream response
stream = client.agents.messages.stream(
    agent_id=agent.id,
    messages=[{"role": "user", "content": "Hello"}],
    stream_tokens=True,
    include_pings=True,  # prevent timeout
)

# Create tool
tool = client.tools.create(source_code="def my_tool(x: str) -> str: ...")
client.agents.tools.attach(agent_id=agent.id, tool_id=tool.id)

# Memory blocks
client.agents.blocks.retrieve(agent_id=agent.id, block_label="persona")
client.agents.blocks.update(agent_id=agent.id, block_label="persona", value="...")

# Folders
folder = client.folders.create(name="docs")
client.folders.files.upload(file=f, folder_id=folder.id)
client.agents.folders.attach(agent_id=agent.id, folder_id=folder.id)

# Conversations (parallel sessions)
conv = client.conversations.create(agent_id=agent.id)
stream = client.conversations.messages.create(conv.id, messages=[...])

# Agent secrets (for tools)
client.agents.update(agent_id=agent.id, secrets={"API_KEY": "..."})

Resources

Platform:

• Letta Cloud (ADE) - Agent Development Environment
• API Keys - Get your API key

Documentation:

• Letta Docs - Full documentation
• Agents Guide - Agent concepts
• Memory Blocks - Memory architecture
• Custom Tools - Tool creation
• Streaming - Real-time responses
• Multi-User - Scaling patterns

SDKs:

• Python SDK - pip install letta-client
• TypeScript SDK - npm install @letta-ai/letta-client

Examples:

• Chatbot Example - Full app example