edoliberty/long-term-memory-in-ai

What it solves

AI systems need a way to store and retrieve information from massive collections of data. This course tackles that by teaching vector search, the core technique behind long-term memory in modern AI. It bridges artificial intelligence and database management, showing how to build systems that can quickly find similar items (such as images, text, or audio) in billions of records.

How it works

The course is a graduate-level seminar (COS 597A at Princeton) taught across 10 lectures, each with detailed PDF notes. It progresses from fundamental concepts like text and image embeddings, through dimensionality reduction and approximate nearest neighbor search, to advanced indexing techniques (clustering, quantization, graph-based methods). Instructors include the CEO of Pinecone, the main developer of FAISS, and the creator of DiskANN. Students apply the material by completing a final project, choosing one of three tracks: theory/research, data science/AI (using Pinecone), or engineering/HPC (extending FAISS). Lectures are supplemented by a comprehensive reading list of seminal papers.

Who it’s for

This is aimed at graduate students (the formal course is at Princeton), but the materials are explicitly offered for free use by any academic, student, or professor. It suits anyone wanting a deep, systematic understanding of vector search and its role in AI systems, especially those interested in building or using vector databases.

Highlights

• Taught by leaders from industry and research: Edo Liberty (Pinecone), Matthijs Douze (Meta/FAISS), and a guest lecture by Harsha Vardhan Simhadri (Microsoft/DiskANN).
• Covers both theory (proofs, asymptotic bounds) and practical systems (FAISS, product quantization, HNSW graphs).
• Includes a large curated bibliography of classic and modern papers on vector search, dimensionality reduction, and quantization.
• Open-contribution model: the repository invites pull requests and issue reports, encouraging community improvement of the course materials.