Edo Liberty

Edo Liberty is the founder and chief scientist of Pinecone, which he started in 2019 to make vector search accessible to engineering teams without the resources of a large technology company. His career spans mathematical research, machine-learning infrastructure at Yahoo and AWS, and building a managed database service. His current focus is long-term memory for AI.

From mathematical research to production systems

Liberty earned a bachelor’s degree in physics and computer science at Tel Aviv University and a doctorate in computer science at Yale, then stayed at Yale as a postdoctoral fellow in applied mathematics. His research includes the 2012 paper Simple and Deterministic Matrix Sketching, which describes how to process a large matrix one row at a time while retaining a much smaller approximation.

The method gives a precise guarantee: for any unit vector, the sketch approximates the squared magnitude of the original matrix’s output, with an error bounded by a chosen tolerance times the matrix’s squared Frobenius norm. That makes the compression useful beyond simply reducing storage. Engineers can reason about what the smaller representation preserves without keeping the entire input matrix. The algorithm is deterministic, and its updates accommodate data arriving as a stream.

Liberty subsequently became a senior research director at Yahoo and head of its New York research lab. The lab built shared machine-learning platforms and worked on advertising, web search, security, recommendations, and email abuse prevention. At AWS, he served as director of research and head of Amazon AI Labs, where researchers and engineers developed algorithms, models, and systems that became parts of services including SageMaker, OpenSearch, Kinesis, and Personalize.

His coauthored Amazon research offers concrete examples of the problems involved:

• Training on large, changing datasets. The SageMaker research examined why industrial-scale machine learning remained difficult and costly, derived requirements for a platform, and described the computational model designed to meet them.
• Working within limited memory and compute. ProxQuant investigated training neural networks with low-precision weights, a useful approach for environments such as mobile devices where resources are constrained.
• Coordinating distributed computation. Research on message scheduling developed algorithms for aggregating gradients across CPUs and GPUs with different connection topologies, treating communication as an optimization problem.

These were collaborative contributions. They show how Liberty’s work extended from mathematical approximation to the practical constraints of training and operating machine-learning systems.

Turning custom vector search into a service

At Yahoo and AWS, Liberty saw AI models and vector search improve applications such as spam detection and recommendations. Those systems depended on custom infrastructure built with substantial engineering and data-science resources. He expected a packaged solution to exist for teams without those resources; finding that it did not led him to found Pinecone.

Pinecone’s founding premise was to provide the storage and retrieval infrastructure needed to build AI applications as a fully managed service. Teams could use vector search without first assembling the expertise and systems required to operate it themselves. The company helped establish vector databases as a distinct category of AI infrastructure.

Liberty led Pinecone as CEO for six years. On September 8, 2025, Pinecone announced that Ash Ashutosh would become CEO and Liberty would become chief scientist. The transition moved Liberty’s responsibilities toward scientific innovation and the company’s expanding AI work.

Retrieval research and long-term AI memory

Liberty’s recent collaborative publications include Accurate and Efficient Metadata Filtering in Pinecone’s Serverless Vector Database, coauthored with Amir Ingber; VQ-bench: A Composable Vector Quantization Framework, with Ashwin Padaki and Ingber; and Nearly Optimal Attention Coresets, with Alexandr Andoni and Eldar Kleiner. The work spans filtering, compact vector representations, and attention approximation, extending his interest in making large-scale computation and retrieval more efficient.

His coauthored writing also covers full-text search architecture, Pinecone Nexus as a knowledge engine for agents, and optimizing Pinecone for agent applications. These topics connect the database’s storage and search foundations to the systems that use retrieved knowledge.

Alongside research and company building, Liberty has taught algorithms in data mining at Tel Aviv University and long-term memory in AI at Princeton. Long-term AI memory is now his stated scientific focus: building systems through which AI can retain and retrieve useful knowledge.