// the find
microsoft/kernel-memory
Research project. A Memory solution for users, teams, and applications.
Kernel Memory is an archived Microsoft research project implementing RAG pipelines as a deployable service or embedded .NET library. It handles document ingestion (PDF, Word, images, etc.), chunking, embedding, and vector search with citations. If you're building a RAG system in .NET and want a reference implementation rather than a greenfield design, this is the most thorough one from Microsoft.
The pipeline architecture is well-thought-out — stateful, step-based ingestion with swappable handlers means you can replace any stage without rewriting everything. Connector coverage is broad: Azure AI Search, Qdrant, Postgres, Redis, Elasticsearch, SQL Server, MongoDB, plus OpenAI/Azure OpenAI/Anthropic/Ollama on the LLM side. The evaluation framework (RAGAS-style metrics: faithfulness, context recall, answer correctness) ships alongside the main code, which most RAG repos skip entirely. The serverless embedded mode and the full web service mode share the same API surface, so you can start local and graduate to a service without rewriting calling code.
It's archived and explicitly unsupported — the caution banner is not marketing hedging, it means no fixes if something breaks against current Azure APIs or newer LLM SDKs. The configuration surface is enormous (appsettings.json with dozens of nested keys) and the setup wizard is a CLI tool that generates JSON, which is fine for demos but painful in real IaC pipelines. The chunking strategy is fixed by default and the options for tuning overlap/size are buried in examples rather than surfaced prominently; naive chunking at default settings will hurt retrieval quality on technical documents. There's no streaming response support — answers come back as completed strings, which matters if you're building any kind of interactive UI on top.