MagiCore
A local-first C# library for building long-term memory into .NET AI agents, assistants, simulations and robots, combining semantic retrieval, an auditable temporal history and spatial memory of timestamped 3D observations.
MagiCore gives .NET applications a memory system for storing what happened, recovering what mattered and recalling it in the right context, running in-process by default with no telemetry and no required hosted service. Production integrations use the standard Microsoft.Extensions.AI and Microsoft.Extensions.VectorData abstractions, so the choice of model and storage provider stays at the application boundary.
The semantic plane offers dense and hybrid retrieval with reranking, entities, relations and multimodal memories; the temporal plane gives audit history, point-in-time reads, rollback and event-time filtering; and the spatial plane holds timestamped 3D observations, reconstructed object beliefs and measured action episodes for robotics use cases. A zero-configuration MemoryService uses a deterministic local store, extractor and lexical embedding generator, and dense or hybrid retrieval combines vector search with BM25 keyword scoring and LLM, Cohere or cross-encoder reranking. Cognitive behaviors include background consolidation of repeated facts into long-term insights, spontaneous associations, and first-person identity-shaped recall.
Built with C# for .NET, distributed as the MagiCore NuGet package, with support for pgvector and Qdrant as vector stores and OpenAI, Ollama or local ONNX models. Apache-2.0 licensed.