Letta (formerly MemGPT) is an open-source framework for building stateful AI agents with persistent long-term memory. It gives LLMs the ability to remember past interactions, learn new skills, and improve through experience. Agents manage their own context windows using a tiered memory system inspired by computer architecture. Letta is model-agnostic, ships with Python and TypeScript SDKs, and offers both self-hosted and cloud deployment options. It is built for developers creating assistants, support bots, coding agents, and any application where AI needs to retain knowledge across sessions.

• Persistent Long-Term Memory: Agents remember context across sessions using a three-tier memory system (core, recall, and archival) modeled after computer memory architecture. No more starting from scratch every conversation.
• Agentic Context Engineering: Agents manage their own context windows by reading and writing to structured memory blocks. Developers set the rules, and agents decide what to remember and what to forget.
• Model-Agnostic Design: Works with OpenAI, Anthropic, vLLM, Ollama, and other LLM providers. Switch models without rebuilding your agent.
• Skill Learning: Agents dynamically learn new skills through experience and can build, version, and share their own tool capabilities over time.
• Agent Development Environment (ADE): A visual interface for creating, testing, debugging, and observing agents. Inspect memory state, replay interactions, and fine-tune behavior without writing code.
• Letta Code: A memory-first coding agent available as a desktop app for macOS, Windows, and Linux. Ranked #1 model-agnostic open-source agent on the Terminal-Bench coding benchmark.
• Sleep-Time Compute: Agents process and organize their raw context offline between interactions, improving response quality without adding latency during conversations.
• Full API and SDKs: Production-ready REST API with Python and TypeScript SDKs. Scale from prototype to millions of agents on the same stack.

• Memory that Actually Works: The tiered memory system (core, recall, archival) provides structured, queryable long-term memory that feels designed for real products, not demos.
• True Open Source: Licensed under Apache 2.0 with self-hosting support. You keep full control over your data and infrastructure without vendor lock-in.
• Active Open-Source Community: Backed by a growing community, Discord channels, forums, and over 100 contributors on GitHub.
• White-Box Transparency: Developers can inspect, edit, and manage every aspect of agent memory and behavior. Nothing is hidden behind a black box.
• Flexible Deployment: Run self-hosted for full control, use the managed cloud for convenience, or bring your own API keys on any plan.

• Steep Learning Curve: Not plug-and-play. You need to design your own memory schema and retrieval strategy, which takes time to get right.
• Performance Overhead: Every memory self-edit breaks prompt caching, which adds latency and increases token costs during complex interactions.
• Tightly Coupled Architecture: The memory system is deeply integrated with the agent runtime. You cannot extract Letta's memory layer and plug it into a different agent framework.
• Higher Token Usage: Memory management can increase model usage and operational costs.
• Complex Agent Design: Long-term memory introduces additional architecture decisions compared to simple chatbots.
• Local Model Performance Varies: Smaller open-source models may struggle with advanced memory workflows.

• Customer Support Agents: Build support bots that remember each customer's history, preferences, and past issues across every interaction. Agents accumulate knowledge over weeks and months, reducing repeat questions and improving resolution times.
• AI Coding Assistants: Use Letta Code to create a personalized coding agent that learns your codebase, remembers your coding patterns, and builds reusable skills. It stores context in git-backed memory and works across any model provider.
• Recommendation Engines: Deploy agents that learn individual user preferences over time to deliver increasingly accurate recommendations. Built Rewards uses Letta agents to build behavioral profiles that improve with each interaction.
• Research and Knowledge Agents: Create deep research agents that build persistent knowledge bases across extended sessions. Instead of starting fresh each query, agents accumulate findings and make connections across prior research.
• Interactive AI Characters: Build CharacterAI-style applications where each character maintains a consistent personality, remembers past conversations, and develops relationships with users over time.
• Enterprise Internal Assistants: Deploy private, self-hosted AI assistants for specialized domains like medical, financial, or legal support. Agents learn internal knowledge bases and maintain full data ownership within your infrastructure.