Frequently Asked Questions

Mnemosyne is a zero-dependency, sub-millisecond memory system for AI agents. It stores and retrieves memories using SQLite with built-in vector search (sqlite-vec) and full-text search (FTS5). No servers, no cloud, no API keys. Just a Python library you pip install.

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.

Yes. While Mnemosyne is a first-class Hermes Agent plugin, it works standalone with any Python application. Import the client, call remember() and recall(), and you have instant memory. There's also an MCP server for LLM tool-calling integration.

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, and Hindsight. 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 environments with Hermes Agent. 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 plugin-based architecture is stable and well-documented.

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 solo developer. Every contribution, no matter how small, makes a difference. You can also support the project personally through Ko-fi at ko-fi.com/axdsan or send Bitcoin directly to 37dTGqu3qFDsptsBPpGH66KsXGdLD5japx. Every bit helps keep this project alive. See the support page for more ways to help: https://mnemosyne.site/support.

Short answer: it replaces them for Hermes Agent users. Mnemosyne is a drop-in replacement for Hermes' built-in memory provider. It handles everything within Hermes' native plugin system: vector search, full-text search, temporal triples, auto-consolidation, the BEAM architecture. You don't need any other memory tool alongside it. For Hermes users running mem0 or Honcho across multiple frameworks and not just Hermes, they serve a different purpose. Mnemosyne fits alongside them, handling the Hermes layer specifically, with deeper integration that general-purpose tools cannot match. But for most Hermes users, Mnemosyne alone is all you need.