Download Mnemosyne OS
Mnemosyne OS stores your digital life in encrypted vaults on your own hardware, treating memory as the foundation rather than an add-on. Documents, images, notes, and agent transcripts become searchable through hybrid semantic and lexical retrieval that operates without an internet connection. The system connects to multiple AI providers or runs local models, while embeddings and ranking stay on your machine. A spatial canvas hosts chat, vaults, a neural map, voice tools, and media utilities across several desktops. MnemoHub extends functionality through signed cartridges, and an MCP server lets external AI clients share the same memory store. A one-time licence unlocks local neural voices, heavy OCR, and premium cartridges.
This desktop environment organizes knowledge by life domain, with each vault maintaining separate protection and consent boundaries. Retrieval adapts to the model you run, scaling from laptop LLMs to cloud systems. A dream state engine consolidates memories during idle periods, resolving contradictions and linking sessions without replacing raw data. Voice input and output work locally or through cloud services with CPU fallback. Theia adds local image embedding and recall through semantic, palette, and category channels. The sovereign wallet manages licensing and credits without accounts or passwords. The interface places widgets on a two-dimensional canvas where position carries meaning.
Mnemosyne OS runs on Windows, macOS, and Linux, offering a perpetual licence for local neural voices, heavy local OCR, and premium MnemoHub cartridges. The MCP server, SDK, and cartridge SDK are published under the MIT licence. Development by XPACEGEMS LLC centers on independent engines for embedding, retrieval, classification, consolidation, image memory, and voice, all connected through validated IPC channels to the user interface. Documentation and technical whitepapers describe the architecture, governance, and design decisions behind the system. The platform positions memory as the central primitive, with AI serving as one of several methods for querying and extending stored knowledge.
| Watches selected folders and ingests content automatically | Keep documents and notes continuously indexed |
| Partitions memory into isolated vaults with consent rules | Separate code, research and journal knowledge pools |
| Combines quantized vector search with exact term matching | Retrieve memories by concept and keyword |
| Embeds via cloud, ONNX or Ollama provider chain | Prevent failed embeddings from becoming invisible |
| Classifies memories by semantic spine and tags | Group and filter content without code changes |
| Consolidates memories during idle rest phases | Resolve contradictions and link past sessions |
| Scales context injection to model tier and mode | Avoid flooding prompts with excessive candidates |
| Accepts text, voice and file input with live retrieval | Converse grounded in previously indexed material |
| Runs independent STT and TTS engines with GPU fallback | Enable voice interaction on varied hardware |
| Embeds vault images locally using SigLIP 2 | Recall visual content through semantic channels |