Mnemosyne-Neural-OS (abbreviated as Mnemosyne OS) is a "sovereign AI operating system," whose core is to give AI memory across sessions and across time, with these memories running entirely on the user's own machine and under the user's control. It is released by XPACEGEMS LLC and employs multiple independent engines rather than a single black-box AI, and it claims not to be in the same category as cloud-based agent memory stores such as Mem0, Zep, and Letta—the latter are memory layers connected to a cloud stack, whereas Mnemosyne OS is a control interface for memory, agents, and models, and runs locally.
Purpose and Positioning The project is described as "everyone is building intelligence, Mnemosyne OS builds relationships": that is, letting AI truly understand the user's memory layer. It manages the resources of AI agents, similar to how Linux manages CPU, memory, disk, and network for programs. Its analogy cites the "LLM operating system" idea of MemGPT (Packer et al., UC Berkeley, 2023, arXiv:2310.08560), but its positioning goes further: it is not a single-session context paging technique, but a system that runs continuously, isolates multiple agents, and persists on the machine as a daemon, rather than a library that can be imported and used immediately.
Main Capabilities and Engines Embedding Engine: A priority-ordered chain of embedding providers (cloud, local ONNX, Ollama), tried one by one, with explicit errors on failure rather than returning empty vectors. Retrieval Engine: An in-memory decrypted vector cache (int8 quantized for scaling), with ANN search and exact term matching merged before final reranking. Spine Engine: Classifies each memory by semantic nature (spine + tags); the taxonomy exists as data rather than hardcoded logic, so new categories can be added without changing code. Dream State: Dual-speed consolidation. A fast, low-latency layer extracts facts during active use; a heavier layer runs when idle/at night, resolving contradictions and linking memories across sessions. Output is appended alongside the original retrieval results and does not silently replace them. Adaptive RAG (the "gearbox"): Context selection (top-k / MMR / low-discrepancy sampling) scales with the model tier being run and the selected thinking mode, rather than injecting all retrieval candidates. Theia Vision Engine: A complete image memory engine; images are 100% locally embedded (SigLIP 2, isolated sidecar) into a separate vector space, with ranking fused through three channels: semantics, pixel color palette, and emergent categories discovered by the engine itself, in chatRecall in the center via thumbnails, and browse in a dynamic gallery. Humans always take priority over the model: rate, pin, describe, teach it, rename, or merge categories. During cold start or while still indexing, it will state this truthfully rather than fabricating "no match." Off by default, placed behind a settings toggle. Speech engine (STT/TTS, fully independent): speech-to-text runs as a small in-process model, while the large model runs in an isolated GPU/CPU sidecar; text-to-speech can use system, cloud, or local (offline binary or GPU voice cloning), scheduled sample-accurately for seamless playback. Without an NVIDIA GPU, it automatically falls back to CPU and will not hard-block. 400 Zod-validated IPC channels: connect the above capabilities to the UI, automatically generated on every build and checked by drift tests.
Usage and ecosystem The flagship application is Mnemosyne OS Infinity Edition, a local-first desktop operating system that runs the sovereign memory core under strict user control; it can run LLMs locally or in the cloud, with the encrypted vault kept on the local machine; inter-agent sync can be transported via libp2p (@mnemosyne-workspace/mnemosync-p2p). Two doors to the IDE:
- Memory-reading agent: DocWatch monitors user-specified folders (such as .claude/.../memory), making content written by the agent searchable alongside documents and code, available since v1.4.3.
- Agent-reading memory: @mnemosyne_os/mcp is a Model Context Protocol server. Add one entry to the client configuration and the agent can search the vault, ask questions in natural language, and write conclusions back. Three of its tools require no application, vault, or token, and directly read the transcripts already written to disk by the runtime framework. Development and build: The open layer (/packages) is the MIT-licensed Layer-2 SDK, the integration interface that can connect applications, build skins, scaffold projects, and evaluate against the Gateway; the MnemoForge CLI (/cli) is a standalone developer tool with its own documentation. To give coding agents memory, use the MCP server. The public Gateway contract lets applications, agents, and skins communicate with the private AI memory runtime, while the cognitive core remains encapsulated and never exposed. All SDK connections use short-lived JWT authentication, listen only on 127.0.0.1, and are governed by the application manifestDeclared permission scope limitations. Cartridge applications in MnemoHub ship with a genuinely readable src/ (React + SDK); MIT cartridges may be freely forked and published; the Cartridge License is source-available, conditioned on running within the Mnemosyne OS ecosystem.
Benchmarks and auditability On the held-out questions of LongMemEval-M, the hybrid mode scores 77.1% (strict review 72.9%). This benchmark uses the full haystack variant, surrounding each question's evidence with approximately 480 distractor sessions. The protocol was published before the run, and the held-out questions had not been answered before; embeddings and search run locally, and answers are generated by the cloud model gemini-3.8-flash. An external audit in September 2025 (Julien Gelee) corrected an erroneous judgment in the August results; after the erratum, strict review was 35/48 (72.9%) and lenient review 37/48 (77.1%). The July 72.9% result is archived with a DOI. Because the reader, engine, and review changed across runs, the figures are juxtaposed but do not constitute progression. The published grader and per-question judgments are officially provided, allowing every score to be rederived with a single command under no-engine, no-network conditions, with the full methodology, root-cause analysis, and raw logs from both runs made public.
Languages and platforms The interface is fully localized into seven languages (English, French, Spanish, German, Portuguese, Russian, Chinese): onboarding, settings, chat, voice assistant, and all dialogs. Switching languages even reselects the embedding model recommended for that language, so retrieval quality follows the language. Supports Windows, macOS, Linux.
Licensing and ownership Running Mnemosyne OS is unlocked by "Engramm," a name taken from "engram" (the physical trace memory leaves in the brain). It is not an account plus monthly subscription, but an on-chain (Base) license bound to a wallet rather than a machine or email; once users hold it, they own a copy and can take it to any device, and anyone can verify it. The same Engramm also powers sovereign pseudonyms and cloud credits. Holding and checking are free, with no gas fees and no need to manage crypto assets: the chain is plumbing rather than a paywall, and users never pay network fees, hold tokens, or approve transactions.
Limitations and information not provided Official materials do not state specific prices, subscription fees, cloud credit consumption rules, or details of commercial terms; hardware minimum requirements for each engine are not provided;a complete compatibility list for MCP with specific clients (beyond Claude) is not provided; benchmark performance beyond LongMemEval-M is not provided; mobile support is not stated. Installation and verification information is subject to the official repository and documentation site.
