New opportunity

Machine Learning Engineer, Memory (R27393)

Epic Games · Cary, United States

About this role

Epic Games is seeking a Machine Learning Engineer, Memory for its Engineering R&D team in Cary, United States. The role focuses on helping foundation models use memory across long-running tasks and different modalities: representing experience, deciding what to retain or discard, and storing, retrieving and exploiting it efficiently. The engineer will work with researchers on projects across the company to develop solutions that extend beyond large context windows and retrieval systems alone.

Responsibilities

Design memory architectures that work across problems, modalities and products.
Develop ways to measure and evaluate memory performance.
Create and deliver production-ready, scalable, high-quality learning models and associated algorithms.
Assess model effectiveness and recommend directions for the ongoing roadmap.

Qualifications

A PhD in Computer Science, Mathematics or a related field, or 3+ years of relevant industry experience.
Experience creating machine learning algorithms for vision and/or language problems and deploying them as production-level systems.
Expert hands-on knowledge of foundation models for vision and/or language, including local deployment and post-training of open-weight models.
Expertise in memory, state or long-horizon behavior in foundation models, potentially gained through language agents and long-context methods, reinforcement learning under partial observability, world models or sequence modeling.
Proficiency with Python and associated machine learning tools and frameworks, including NumPy, SciPy, scikit-learn and PyTorch; experience collaborating with developers using a flexible, agile approach.

Benefits listed include employer-paid medical, dental, vision HRA, long-term disability and life insurance premiums for employees and dependents; a 401(k) with match; mental well-being support; paid company-wide breaks; unlimited PTO and sick time; and a paid sabbatical after seven years of employment.

For full details about this role, click “Apply on company site” to view the employer’s original posting.

Skills for this role

Machine LearningFoundation ModelsMemory ArchitecturesComputer VisionNatural Language ProcessingModel Post-TrainingModel DeploymentModel EvaluationReinforcement LearningWorld ModelsSequence ModelingLanguage AgentsLong-Context MethodsPythonNumPySciPyscikit-learnPyTorchAgile Development

Checking your profile…

Preparation suggestions, not the employer’s actual interview questions. Always check the original posting for current requirements.

Understand the role

What matters most

Design general memory architectures
The role calls for architectures that work across problems, modalities and products, enabling foundation models to represent experience and store, retrieve and use it. Prepare to explain architectural choices rather than relying on longer context windows or retrieval alone.
Measure memory performance
The posting explicitly asks for new ways to measure and evaluate memory. Prepare to describe how you would test whether retained experience improves behavior over long-running tasks, including comparisons against simpler approaches.
Deliver production-ready models
The engineer will help deliver scalable, high-quality learning models and algorithms. Prepare to discuss how a vision or language ML approach moves from experimentation to a production-level system.
Apply hands-on foundation-model expertise
The qualifications call for local deployment and post-training of open-weight vision and/or language models, plus expertise in memory, state or long-horizon behavior. Prepare a technically precise account of relevant model and memory design decisions.
Assess results with researchers and developers
The role involves researchers on projects across the company, critical assessment of model effectiveness and roadmap recommendations. Prepare to explain how evidence would inform a change in technical direction while collaborating with developers.

Likely challenges to think through

Deciding what memory deserves retention
Likely preparation challenge, not a confirmed employer problem: keeping too much experience can undermine efficiency, while discarding too much can lose information needed later. Prepare to compare retention policies and explain how you would test their effects on later behavior.
Making evaluation credible over long horizons
Likely preparation challenge: an apparent memory gain may come from a larger context or better retrieval rather than useful stored experience. Prepare to propose controlled comparisons and tests that reveal when memory helps, fails or becomes stale.
Generalizing without sacrificing deployability
Likely preparation challenge: an architecture designed for several modalities and products may be harder to keep efficient and production-ready. Prepare to discuss where you would share components, where you would allow specialization, and how you would validate the resulting trade-offs.

Practice your answers

Technical practice

How would you design a memory architecture that transfers across problems and modalities?

What to coverThe role calls for memory architectures that work across problems, modalities and products. Suggested preparation: describe how you would represent experience, separate reusable components from modality-specific ones, and test whether the design generalizes.

How should a long-running model decide what to remember or discard?

What to coverThe posting emphasizes deciding what experience is worth keeping. Suggested preparation: discuss criteria such as future usefulness, redundancy and storage cost, including how an incorrect discard decision could affect later behavior.

How would a model store, retrieve and use memories efficiently?

What to coverEfficient storage, retrieval and exploitation of memory are stated goals. Suggested preparation: explain how information would move from new experience into memory and back into model behavior, and where efficiency might conflict with usefulness.

How would you measure whether a model's memory is working?

What to coverDefining new ways to measure and evaluate memory performance is an explicit duty. Suggested preparation: propose long-horizon tasks and measures that distinguish retaining information, retrieving it when needed and using it to improve an outcome.

How would you compare a memory architecture with large context windows or retrieval alone?

What to coverThe page states that large context windows and retrieval systems are not enough for the work envisioned. Suggested preparation: describe controlled comparisons that isolate the benefit of persistent or updated state, while accounting for computational cost.

What does partial observability imply for a model's internal state?

What to coverThe posting names reinforcement learning under partial observability as one possible source of relevant experience. Suggested preparation: explain why past observations may matter when the current input is incomplete and how a compact state might help with long-horizon decisions.

How would you post-train and locally evaluate an open-weight foundation model for a memory task?

What to coverHands-on knowledge of local deployment and post-training of open-weight vision and/or language models is requested. Suggested preparation: outline a reproducible baseline, a post-training objective tied to memory behavior, and checks for improvement beyond the training examples.

How would you move a memory-enabled learning model toward production?

What to coverThe role includes delivering scalable, high-quality learning models and algorithms, and asks for production-level deployment experience. Suggested preparation: discuss validation, reliability and resource trade-offs when moving from an experiment to a deployed system.

How would you diagnose a memory model that performs well on short tasks but poorly on long ones?

What to coverLong-horizon behavior and critical assessment of model effectiveness are central to the posting. Suggested preparation: separate possible failures in representation, retention, retrieval and use of memory, then explain what evidence would distinguish them.

How would you use experimental results to recommend a research direction?

What to coverThe engineer is expected to assess model effectiveness and make recommendations for the ongoing roadmap. Suggested preparation: explain how you would weigh evaluation results, limitations across modalities and production feasibility before recommending further work or a change of direction.

Behavioural practice

Tell me about a time you collaborated with researchers on an uncertain technical problem.

What to coverThe role involves working with researchers on projects across the company. Suggested preparation: use a real situation/action/result example showing how you agreed on a testable question, contributed to the work and handled differing interpretations of the results.

Describe a time you adapted your approach while working with other developers.

What to coverThe posting asks for collaborative work with developers using a flexible, agile approach. Suggested preparation: give a real situation/action/result example that shows what changed, how you coordinated the adjustment and what it achieved.

Tell me about a time you turned an ML experiment into a production-level system.

What to coverThe posting seeks experience deploying vision and/or language ML algorithms and helping deliver production-ready models. Suggested preparation: use a real situation/action/result example that distinguishes your contribution from the team's and explains a quality or scalability trade-off.

Describe a time your evaluation challenged a preferred technical direction.

What to coverCritically assessing model effectiveness and recommending roadmap directions are stated responsibilities. Suggested preparation: give a real situation/action/result example covering the evidence, how you communicated it and what decision or next experiment followed.

Sharpen your knowledge

Topics to prepare

Memory architecture across modalities
The role calls for architectures that represent experience and decide what to retain, discard, store, retrieve and use across problems, modalities and products. Suggested exercise: sketch one architecture for a long-running vision or language task and explain how each memory operation would change for another modality.
Long-horizon state and partial observability
The posting seeks expertise in memory, state or long-horizon behavior; it names language agents, long-context methods, reinforcement learning under partial observability, world models and sequence modeling as possible backgrounds. Suggested exercise: compare two of these approaches on what state they preserve, how they update it and where they may fail.
Memory evaluation
Defining ways to measure memory performance is an explicit responsibility. Suggested exercise: design a test in which useful information appears early in a long task, then specify measures for later recall, effective use of that information and the cost of retaining it.
Effectiveness and roadmap trade-offs
The engineer will assess model effectiveness and recommend roadmap directions. Suggested exercise: prepare a short decision memo comparing a proposed memory approach with a long-context or retrieval-only baseline, stating what evidence would justify further work.

Skills to revise

Implementing a memory mechanism
The role involves creating learning models and associated algorithms. Suggested exercise: build a small PyTorch model that updates a bounded memory over a sequence; demonstrate how it selects information to keep and uses it at a later step.
Local deployment and post-training
Hands-on knowledge of locally deploying and post-training open-weight foundation models for vision and/or language is explicitly requested. Suggested exercise: rehearse a local deployment and a small post-training experiment with an available open-weight model, then explain the setup, observed change and limitations.
Evaluation in Python
The posting names Python, NumPy, SciPy, scikit-learn and PyTorch. Suggested exercise: write a reproducible comparison of memory variants, calculate task results and resource costs, and inspect cases where retained information did not help.
Production-level model delivery
The role seeks experience deploying vision and/or language algorithms as production-level systems and delivering scalable, high-quality models. Suggested exercise: review a model deployment you can discuss and prepare a technical walkthrough of its inference path, scaling constraints, quality checks and failure handling.
Collaborative technical iteration
The posting calls for collaboration with researchers and developers using a flexible, agile approach. Suggested exercise: practice presenting an experiment result with a clear recommendation, unresolved risks and a concrete next iteration for researchers and engineering partners.

Your preparation checklist

Select a relevant project walkthrough
The posting calls for production-level vision or language ML algorithms. Prepare one example, if available, showing the problem, your contribution, deployment decisions, results and trade-offs; otherwise, prepare a clearly labeled design example.
Sketch a long-horizon memory architecture
As a preparation exercise, diagram how a foundation model could represent experience, decide what to retain or discard, and store, retrieve and use memory across tasks or modalities. Explain where a large context window or retrieval alone would fall short.
Draft a memory evaluation plan
The role includes defining memory measures. Prepare a small test plan for long-running tasks, with proposed metrics, baselines, failure cases and a way to distinguish useful memory from unnecessary retention.
Prepare a hands-on model example
The posting seeks local deployment and post-training experience with open-weight foundation models. If you have relevant work, outline the model, Python/PyTorch tooling, post-training method and observed results; do not present a hypothetical exercise as past experience.
Work through a production trade-off
Prepare to discuss how you would make a memory-enabled model scalable and high-quality. A short design note could cover storage and retrieval costs, latency, reliability, and how you would assess effectiveness before recommending a roadmap direction.
Form questions about the research-to-product path
Bring questions about which modalities and long-horizon problems matter most, how memory performance is currently evaluated, and how researchers and developers collaborate across projects. These are questions to ask, not established company processes.

Availability last checked: Nov 4, 2026. Status can change after a check; confirm on the employer’s website. Report an issue · Our checking policy

KEEP EXPLORING

Similar AI jobs

Related skills and specializations in United States. Matched to this role, not your profile.

Explore more jobs

Finding similar opportunities…