Letta

Letta (formerly MemGPT) is a stateful agent platform built around the idea that AI agents should manage their own memory like an operating system manages virtual memory. The project gained attention as MemGPT — a research paper demonstrating that LLMs could be given explicit memory management tools (read from archival memory, write to archival memory, search core memory) and would learn to use them effectively. Letta is the production platform that evolved from that research.

The core innovation is treating the LLM's context window like main memory in a computer. The agent has 'core memory' (always in context — like RAM), 'recall memory' (searchable conversation history — like a page file), and 'archival memory' (long-term storage — like a hard drive). The agent itself decides when to page information in and out, search its archives, or update its core memory blocks. This self-directed memory management means the agent adapts its memory strategy to the conversation rather than relying on fixed retrieval logic.

Letta agents are persistent and stateful. They run as server processes with their own memory, conversation history, and tool access. You interact with agents through a REST API, and the agent maintains state between calls. This is different from most agent frameworks where agents are stateless functions that reconstruct context on each invocation.

The platform includes an agent development environment (ADE) with a visual interface for creating agents, defining memory blocks, attaching tools, and testing interactions. Agents can be deployed as API endpoints and connected to external tools, data sources, and other agents.

Letta supports multi-agent architectures where agents can communicate with each other, share memory, and collaborate on tasks. Each agent maintains its own state and memory, creating a genuinely multi-agent system rather than a single LLM with multiple personas.

The tradeoffs: Letta's self-directed memory management is powerful but can be unpredictable. The agent decides when to search or update memory, which means it sometimes doesn't retrieve relevant information or updates memory unnecessarily. The server-based architecture adds operational complexity compared to simpler agent frameworks. And the transition from academic research project to production platform is still ongoing — some features are polished while others feel experimental.

For applications that need truly persistent, stateful agents with sophisticated memory management — personal assistants, long-running customer relationships, complex multi-session workflows — Letta offers capabilities that simpler frameworks can't match.