Letta

Letta is an open-source framework for stateful agents, created by the Berkeley researchers behind the MemGPT paper, Charles Packer and Sarah Wooders. It treats memory as a managed resource: an agent edits its own context, moves information between in-context and external memory, and so carries knowledge across sessions past the model's context window. A server-based architecture exposes agents through a REST API with persistence handled for the developer. Treating memory as a managed resource lets agents accumulate knowledge over time, which suits assistants that persist across many sessions. The development environment exposes an agent's memory and reasoning so teams can inspect what it retained.

Key Capabilities:

• Self-editing memory that persists across sessions

• Memory management beyond the model's context window

• Agents exposed as services through a REST API

• The Agent Development Environment for visual inspection of state

• Model-agnostic operation across commercial and open LLMs

• Apache 2.0 license with Python and TypeScript SDKs