Mnemosyne
Local Memory for LLM Conversations
Grounded recall of what you have actually said and decided.
Mnemosyne is a locally hosted memory system for your LLM conversations. It ingests your chat archives and documents, keeps them searchable, and gives a chat model grounded recall, running entirely on your own machine against your own local models.
It is shared between your LLM clients, not between people. Each person runs their own install against their own database, and nothing is sent anywhere.
What it does
Import
ChatGPT, Claude, Gemini, Perplexity, and ChatLLM archives, plus PDFs, Office documents, text, and images, with OCR done locally.
Retrieve
Hybrid vector and full-text search, fused and re-scored by a cross-encoder reranker, which lifted document recall@10 from 0.567 to 0.742.
Chat
Chat against your own corpus in a web dashboard, with the passages the answer used shown alongside it.
Serve
The same memory is available to Claude Desktop, Claude Code, and other MCP clients over stdio or HTTP.
Brief
Summaries of what you concluded about a topic. A machine-inferred conclusion never reaches the model as fact until you confirm it.
Back up
Back up and restore the whole memory as one encrypted file, directly from the dashboard.
Local by design
Every service binds to your own machine only; nothing is exposed to your network. Embeddings, reranking, OCR, and chat all run against local models, so your conversations and documents never leave the machine they live on.
Status
- Status
- Built and running: import, hybrid retrieval with reranking, dashboard chat, MCP serving, briefs, and encrypted backup.
- License
- Apache 2.0