Kernel Memory
An archived Microsoft research project demonstrating a multi-modal memory service for LLM applications, indexing documents into embeddings for retrieval-augmented generation with citations.
Kernel Memory is a reference implementation, not a supported product, showing how to build a Memory service for LLM applications. It indexes datasets through a continuous data pipeline and supports retrieval-augmented generation, synthetic memory, prompt engineering and custom semantic memory processing.
A default ingestion pipeline extracts text from documents, partitions it into chunks, generates embeddings with an LLM embedding model, and saves the embeddings into a vector index such as Azure AI Search or Qdrant, so that natural-language queries can retrieve answers with citations back to source documents. Document ownership can be attached so private information can be safeguarded.
It is available as a web service, a Docker container, a plugin for ChatGPT, Microsoft Copilot and Semantic Kernel, and as an embeddable .NET library, and can be deployed to Azure or run via .NET Aspire.