mnemostack
A self-hosted, durable memory and retrieval layer for AI agents and apps, combining vector, keyword, temporal and graph recall over your own Qdrant and Memgraph.
mnemostack fuses semantic vector search, BM25 keyword search, temporal recall and graph recall with Reciprocal Rank Fusion and an eight-stage ranking pipeline, adding payload filters for multi-tenant isolation, optional LLM answer synthesis with confidence scores and citations, and an ingest path that enriches and projects structured fields. It is used through one recall(query) call, as a Python library, an HTTP service, or an MCP server.
Its flagship use is durable memory for long-running agents, so decisions and context survive context compaction and session restarts instead of paying a re-orientation cost on every run. The same engine also supports RAG over mixed corpora and time-aware search.
It runs over a self-hosted Qdrant vector store, with Memgraph optional for graph-backed memory, and installs with pip install 'mnemostack[mcp]'.