Research Scientist

Engram Lab•San Francisco, CA

5d•Onsite

About The Position

Engram's mission is to make AI understand personal context. Current AI models are powerful but lack personal knowledge, requiring them to re-read files and burning tokens for basic questions. Engram trains models to study user context and anticipate questions by forming compact memories called 'engrams'. This approach scales by improving inference time with more training time context. Engram is collaborating with major AI companies like Microsoft, Notion, and Harvey, and has secured $98M in funding from prominent investors. The company aims to shift AI's focus from global knowledge to individual user context.

Requirements

A deep background in machine learning, with strong fundamentals in inference serving systems, KV cache design, or latency-sensitive model deployment.
A track record of rigorous ML research — publications, strong open-source contributions, or equivalent demonstrated depth.
Extensive experience in at least one area directly relevant to our work: continual learning, memory architectures, test-time training (TTT), parameter-efficient finetuning, context compression, retrieval, synthetic data, distillation, or agents.
Comfort working up and down the stack — you understand both the research question and the system it runs on — not just describe an idea in a paper and hand it off.
Strong technical communication: you can explain complex ideas simply and engage in high-bandwidth, generative technical conversation.

Nice To Haves

Experience bridging research and product — shipping things that real users interact with.
Familiarity with LLM training infrastructure.

Responsibilities

Design experiments
Develop new recipes
Build evals
Shape the product
Design and evaluate methods for encoding large, heterogeneous document corpora into compact parametric memory (e.g., LoRA/adapter-based representations, prefix tuning, state-space methods)
Understand what makes synthetic training data generalize, and develop self-study pipelines that allow models to reflect on and consolidate new context
Tackle catastrophic forgetting, sequential updates, knowledge conflicts, and the tradeoffs between in-weights memory and agentic retrieval
Explore reinforcement learning methods that let models improve from interaction and feedback in real deployment settings
Empirically study how model capacity, data scale, and compute interact; develop the scaling laws that inform our product roadmap