Hybrid Research Engineer / Scientist, Self-Improvement na Letta

Building the AI Operating System

We believe AI's future isn't just more intelligent language models, it's agents that remember, learn, and evolve over time. We're building the AI Operating System to turn stateless models into perpetual and self-improving intelligence.

Letta is founded by AI researchers from UC Berkeley's Sky Computing Lab (the same lab that produced Spark → Databricks and Ray → Anyscale), and the creators of MemGPT. We’re backed by leaders in AI and infra, including Jeff Dean (Chief Scientist at Google AI), Clem Delangue (CEO of HuggingFace), and Cris Velenzuela (CEO of Runway). Our technology is already being deployed in production to create autonomous, self-improving AI agents at companies like 11x and BILT Rewards.

We’re looking for incredibly talented individuals to join our team of AI researchers and engineers to create the first AI OS with human-like memory and learning capabilities. Join us and build the future of open AI, in the heart of San Francisco.

👾 Read more about Letta on TechCrunch, our blog, and on our GitHub (17k+ stars).

📃 Research whitepaper: Sleep-time Compute: Beyond Inference Scaling at Test-time

Note that this role is in-person (no hybrid), 5 days a week in downtown San Francisco.

Your role

Our goal as a company is to empower developers to build state-of-the-art LLM agents to power their own applications. To achieve this, we need to build and maintain a strong lead in our underlying agent technology. As a Research Scientist, your primary responsibility will be to design and execute an R&D strategy to ensure that the company’s core LLM agent technology remains state-of-the-art. Simply put, your mission is to push the limits of what is possible with LLMs.

Responsibilities:

• R&D of the core agent framework, including context management techniques used to prompt the LLMs within the agent framework

  • Prototyping new context management and memory systems (e.g. better state→prompt compilers, more complex memory systems, multi-threading of multiple LLM processes, improved planning for multi-step reasoning, etc.)

  • Development of the agent interaction loop, encompassing tool execution, parsing, and more

• R&D of LLM serving methods to improve serving agent workloads (e.g. constrained decoding and prefix caching)

• Model evaluation and finetuning, such as finetuning state-of-the-art open-weights models on agent data traces (which we plan to release publicly as free models on HuggingFace)

Technical reports and blog posts: As our founding research scientist, you will also have significant agency in writing and publishing technical blog posts and whitepapers, which will be an important driver of our marketing and recruiting efforts.

Compute resources: You will be provided with significant resources to accomplish your work, including access to the latest GPUs for training/finetuning via both spot and reserved instances, and a dedicated budget for model serving and API credits across all the frontier model providers.

Signs it could be a great fit:

• You want to be an integral part of turning a tiny startup into a trillion dollar company. You wonder what it would have been like to be at OpenAI when it was just a dozen people, or Google when it was just a couple grad students in a garage.

• If you work / worked at a large tech company: you felt physically pained by the red tape and bureaucracy wedged between you and your potential impact.

• You’re excited to go head-to-head with tech giants, frontier labs, and other startups that are many times our size in both headcount and funding.

• You are anti closed frontier AI that is controlled by a few private tech companies.

Signs it’s a bad fit:

• You like having things planned out far ahead of time, and get stressed out when there’s nobody telling you exactly what to do. We look for people that thrive in ambiguity and can drive their own agenda.

• You want to work a 9-5, and value clear separation of work from life. The stakes are high, and the only moat is execution and velocity. We work hard because incredible outcomes require incredible sacrifice – operating on a strict 9-5 guarantees failure.

• You value titles or want to people-manage. Letta is a flat company where every researcher and engineer is an individual contributor.

• You’re not interested in talking to customers, and prefer to stick to one part of the stack. At Letta everyone on the team engages directly with our customers and works across the stack.

Our Interview Process:

• Initial screen (30 min)

• Technical screen (1-1.5 hours)

• Paid in-person work trial (2 days onsite in SF)