What Letta does
Letta is a stateful AI agent platform that aims to make “memory management” the core primitive for building long-running, model-agnostic agents. Instead of treating LLM usage as stateless request/response, Letta positions an agent harness/runtime as responsible for maintaining persistent identity, context, and long-term experience—so that agents can keep improving over time and can be ported or re-run with different underlying model providers without losing the agent’s stored state.
Product-wise, Letta’s platform centers on “Letta Code” (their memory-first agent runtime) and the surrounding developer surfaces: (1) a model-agnostic hosted offering (“Letta Cloud”), (2) an agent developer environment for building, debugging, and deploying stateful agents, and (3) SDKs/APIs that let developers connect custom clients and tool ecosystems to stateful agents while keeping memory and state under the developer’s control.
A key differentiation Letta emphasizes is transparency/white-box handling of agent memory and reasoning context. In their initial announcement, Letta describes “white-box memory” as making the prompts and memories provided to the LLM on each reasoning step visible to developers, contrasting this with agent frameworks where memory and intermediate reasoning inputs are less inspectable.
Strategically, Letta is using rapid iteration on its memory and harness architecture—plus a growing set of developer interfaces—to move from an “agent framework” model toward an “agent runtime” that is meant to be extensible across models and deployment environments. Examples of recent product direction include launching Letta’s Agents SDK (TypeScript) and pushing toward a runtime that uses git-backed context repositories and skills/subagents to support more complex, long-horizon agent behaviors.