← back
Stateful Agents — Full Workshop with Charles Packer of Letta and MemGPT
Takeaway
Memory/statefulness — not bigger models or more tools — is the binding constraint preventing agents from becoming actually useful in production.
Summary
- Charles Packer (Letta, MemGPT) argues 'agent = LLM in a loop' misses the key point: transformers are stateless, so the loop only works if you have a mechanism to update state — statefulness equals memory.
- Default state-handling is appending to a Python list — fine for 2022-2024 toy demos, fatally inadequate for real production agents.
- MemGPT/Letta builds an LM-OS: long-term memory, working memory, and explicit memory tools so agents form new memories like humans rather than relying solely on context window or weights.
- Workshop format: Docker image + notebook so attendees run a Letta server locally and build stateful agents interactively.
memorylettastateful-agents
Original description
A cornerstone of human intelligence is the ability to learn: as humans interact with the world, we form new memories and can adapt from experience. In this workshop, participants will learn about “stateful agents”: agents that live indefinitely, and can form new memories by learning from data and experience. The workshop will cover best practices for context and state management for agents, as well as how to avoid common challenges with agents (e.g. context overflow errors, memory loss, and lack of user personalization). In this workshop you will: - Learn the principles behind context and memory management from the lead author of the MemGPT paper - Learn about stateful agents in practice and deployment (scaling to hundreds of thousands of agents) - Get hands on experience building stateful agents with the Letta framework and ADE (Agent Development Environment) Notebook materials: https://github.com/letta-ai/tutorials/tree/main/python Charles: https://x.com/charlespacker Letta: https://x.com/letta_ai