Memory-first AI agents that learn from experience
Letta builds an AI operating system for stateful LLM agents with advanced memory management, enabling persistent, self-improving intelligence. Spun out from UC Berkeley's MemGPT project, it serves developers creating agents that remember interactions, adapt over time, and compound improvements without model lock-in. This addresses the core limitation of stateless LLMs by providing infrastructure for long-term learning and evolution.
Designs long-term memory architectures for LLMs, builds multi-type memory systems, researches agent memory sharing and context management, and runs evaluations. Requires deep LLM/retrieval expertise and impactful research track record.
Develops methods for AI agents to self-improve post-training through prompt optimization, continual learning from long-horizon tasks, hypothesis testing, and scalable experiments. Requires expertise in LLMs, agent frameworks, and impactful research track record.
Pioneers post-training techniques to enhance LLMs for agentic systems, focusing on tool-use, continuous updates, synthetic data infrastructure, and capability evaluations. Requires Python/PyTorch proficiency, post-training expertise, and proven research impact.
Builds scalable backend services, APIs, and OSS framework for production AI agent systems. Requires strong Python expertise, service architecture, SQL databases, and AI tooling familiarity; in-person in San Francisco.