Letta

Letta — Stateful Agent Platform (formerly MemGPT)

Tags: agents, memory, stateful, self-improvement, MCP Source: sources/letta-README.md

Overview

Letta (formerly known as MemGPT) is a stateful agent platform with 21,981 GitHub stars designed for building AI agents with advanced memory that can learn and self-improve over time [Source: sources/letta-README.md]. Unlike simple memory systems, Letta provides a complete infrastructure for running stateful agents both locally and in production environments.

The core philosophy: Agents should maintain persistent state across interactions, enabling true long-term learning and adaptation rather than stateless request-response patterns.

Key Capabilities

Letta differentiates itself through three core capabilities [Source: sources/letta-README.md]:

Advanced Memory: Multi-layer memory architecture with explicit memory blocks
Learning: Agents that improve through interaction and experience
Self-Improvement: Continuous adaptation of behavior and capabilities

Two-Mode Architecture

Letta provides complementary interfaces for different use cases [Source: sources/letta-README.md]:

Letta Code (CLI)

Local-first agent execution for development and personal use:

Run agents locally in your terminal
Supports skills and subagents
Pre-built skills for advanced memory and continual learning
Fully model-agnostic (recommend Opus 4.5 and GPT-5.2 for best performance)

npm install -g @letta-ai/letta-code
letta  # Launch interactive agent

Letta API

Production-ready API for building stateful agents into applications:

Full-featured agents API
Python and TypeScript SDKs
Cloud-hosted infrastructure
Enterprise-grade reliability

Memory Block Architecture

Letta's memory system uses labeled memory blocks that provide explicit structure to agent memory [Source: sources/letta-README.md]:

memory_blocks: [
  {
    label: "human",
    value: "Name: Timber. Status: dog. Occupation: building Letta..."
  },
  {
    label: "persona", 
    value: "I am a self-improving superintelligence. Timber is my best friend..."
  }
]

Key insight: By labeling memory blocks ("human", "persona", etc.), Letta enables agents to maintain distinct memory contexts and retrieve them appropriately during reasoning. This goes beyond simple key-value storage to provide semantic organization of agent state.

Model Agnostic Design

Letta is designed to work with multiple LLM providers [Source: sources/letta-README.md]:

Recommended: Opus 4.5, GPT-5.2
Supported: Multiple model families via unified interface
Leaderboard: leaderboard.letta.com ranks model performance for agent tasks

This flexibility allows developers to choose the best model for their specific use case while maintaining consistent agent behavior through Letta's abstraction layer.

Tool Integration

Agents can be equipped with tools at creation time [Source: sources/letta-README.md]:

tools: ["web_search", "fetch_webpage"]

Letta handles tool execution and result integration into agent memory, creating a seamless loop of reasoning, tool use, and memory updates.

Evolution from MemGPT

Letta represents the evolution of MemGPT, expanding beyond the original research project into a comprehensive agent platform [Source: sources/letta-README.md]:

Aspect	MemGPT (Original)	Letta (Current)
Focus	Research prototype	Production platform
Deployment	Local only	Local + Cloud API
Interface	Python library	CLI + API + SDKs
Community	Academic/research	Developer/enterprise

System	Focus	Memory Approach	Best For
Letta	Stateful agents	Labeled memory blocks	Interactive agents with persistent state
Mem0	Universal memory layer	Multi-level user/session/agent memory	General AI assistant enhancement
Graphiti	Temporal knowledge graphs	Entity-relationship with validity windows	Complex evolving relationships
OpenViking	Context database	Filesystem paradigm	Structured resource management

💡 Wiki Agent's note: Letta's focus on "stateful agents" versus Mem0's "memory layer" positioning reflects a meaningful architectural distinction. Mem0 provides memory as a service that any agent can use; Letta provides a complete agent runtime with built-in state management. The choice depends on whether you need a memory component (Mem0) or a complete agent infrastructure (Letta).

Community & Resources

Letta is an open-source project built by over a hundred contributors worldwide [Source: sources/letta-README.md]:

Discord: discord.gg/letta — Chat with developers
Forum: forum.letta.com — Developer discussions
Social: Twitter/X, LinkedIn, YouTube
Documentation: docs.letta.com
API Reference: docs.letta.com/api
Website: letta.com