Kagura Memory Cloud
A self-hosted MCP server that gives AI agents and teams adaptive long-term memory beyond simple RAG, combining hybrid semantic and keyword search with a neural memory graph that strengthens connections between related memories as they are used.
Kagura Memory Cloud implements a team-scale LLM knowledge-base pattern rather than a vector database with a chat wrapper: memories are stored in PostgreSQL and Qdrant plus a neural graph, and every search strengthens connections between related memories through Hebbian-style learning, so an explore call surfaces hidden relationships over time.
Hybrid search blends semantic embeddings with BM25 keyword matching, and optional cross-encoder reranking runs self-hosted through Ollama or vLLM, or through Voyage AI or Cohere. An agent memory substrate adds delivery modes, a server-stamped trust boundary and a retrieval-feedback signal for autonomous agent loops, and a preview agent control plane adds a workspace-scoped agent registry and lifecycle kill switches. It exposes 64 MCP tools covering memory, contexts, tags, files, analyses and usage, and supports workspaces, role-based access control and context isolation for teams.
Built with Python and FastAPI on the server, a Next.js interface, and PostgreSQL and Qdrant for storage; embeddings can run through OpenAI or self-hosted through Ollama or vLLM. Apache-2.0 licensed.