What Letta does
Letta is an AI agent platform and research-to-software company focused on building “stateful” agents—agents whose identity and memory persist across sessions—so they can learn from experience over time rather than operating as stateless, single-session LLM chats. Letta presents its work as an agent runtime/harness for model-agnostic deployment: developers can connect LLMs of choice to a persistent agent system that manages context and long-lived memory. Letta’s flagship direction is Letta Code, which it describes as a model-agnostic agent harness with persistent memory, plus a developer SDK and tooling that support building and deploying agents that retain and evolve their state across model providers and machines. The company originally emerged from MemGPT research and positions itself as evolving beyond “memory as a plugin” toward memory and context management as core responsibilities of the agent harness itself. Letta’s recent product messaging emphasizes (1) a memory-first, inspectable architecture, (2) a shift from server-centric features toward runtime-native and client-side “computer use” workflows, and (3) extensibility via skills (and harness-level extension mechanisms such as Mods) so agent capabilities can be composed and improved over time. Letta targets developers and teams building long-running agent applications (including coding agents), and it also supports production deployment by offering persistent agents that can run across different model endpoints while preserving the agent’s memory and identity.