Frequently Asked Questions

Mnemosyne is a universal memory engine for AI agents. It exposes 30+ memory tools through MCP and a Python SDK, so any agent that speaks MCP — Codex, Claude Code, Cursor, Windsurf, OpenWebUI, OpenClaw, Hermes — can call it the same way. Under the hood, it's zero-dependency, sub-millisecond, and SQLite-backed with native vector (sqlite-vec) and full-text (FTS5) search. One pip install — or no install at all if your client already speaks MCP.

No. Mnemosyne is 100% local. All data lives in a SQLite file on your machine. There are no external API calls, no telemetry, no cloud dependencies. It works offline, in airplane mode, on a plane. Your memory never leaves your computer.

Yes. The core Mnemosyne library is free, open source (MIT), and always will be. No paid tiers, no feature gates, no 'contact sales.' You get the full local-first system with every feature, forever. We are also exploring an optional cloud-hosted solution for those who want managed infrastructure, but the core local version will never become paid or lose features to a paid tier.

Mnemosyne runs anywhere an MCP client runs. Out of the box: Hermes Agent (plugin + MCP), Claude Code, OpenAI Codex CLI, Cursor, Windsurf, OpenWebUI, and OpenClaw. Any other MCP-compatible client works too — point it at `mnemosyne mcp` and the tools register automatically. There's also a Python SDK if you want to embed memory directly in your own application.

Mnemosyne is 43-500x faster because there are no network roundtrips. It has zero dependencies (just Python stdlib + ONNX). Your data stays on your machine. No Docker, no Postgres, no API keys, no rate limits. It also uses a unique three-tier BEAM architecture no other memory system offers.

Yes. Mnemosyne integrates sqlite-vec for native vector similarity search, combined with FTS5 for full-text search and importance scoring. This hybrid ranking (50% vector + 30% FTS + 20% importance) gives you the best recall without needing a separate vector database.

No. Mnemosyne runs entirely on CPU. Vector search, embedding generation, and consolidation all happen locally without GPU acceleration. Benchmarks show sub-millisecond query times on standard hardware.

Yes. Mnemosyne includes built-in importers for Mem0, Zep, Honcho, Hindsight, and Hermes Holographic Memory. Run one command and your existing memories are migrated into Mnemosyne's SQLite store. See the migration docs for details.

Yes. Mnemosyne is used in production across multiple agent frameworks — Hermes, Claude Code, Codex, OpenWebUI, and others. It has nearly 300 tests, a formal BEAM benchmark scoring 65.2% (leading published SOTA), and supports memory at scale from 100K to 1M+ tokens. The MCP interface is stable, the BEAM architecture has been validated against ICLR 2026 baselines, and all 3.x schema changes have been additive — no breaking migrations.

Mnemosyne uses automatic sleep cycles. Old working memories are compressed into episodic summaries, freeing context space while preserving key information. Consolidation can run on a schedule (configurable auto_sleep interval) or on demand via the sleep() API.

Privacy is the core design principle. Every memory stays in your SQLite database on your machine. No data is ever sent to external servers. No telemetry. No analytics. No tracking. You own the file, you control the backup, you decide what happens to your data.

Mnemosyne welcomes contributions. Star the repo, report bugs, suggest features, improve documentation, or submit pull requests. The project is MIT licensed and maintained by a community of developers. Every contribution, no matter how small, makes a difference. You can also support the project financially through our OpenSource Collective on OpenCollective at opencollective.com/mnemosyne, GitHub Sponsors, or Ko-fi. Every bit helps keep this project alive. See the support page for more ways to help: https://mnemosyne.site/support.

Mnemosyne replaces them at the memory layer for every agent and tool it integrates with — Codex, Claude Code, Cursor, Hermes, OpenWebUI, OpenClaw, and any MCP client. You get one SQLite file you control, sub-millisecond recall, and zero vendor lock-in. For multi-framework setups where different agents need to share memory, the sync engine (added in 3.9.0) lets Mnemosyne instances exchange state bidirectionally. But for any single agent, Mnemosyne alone is enough.