Launching today
Mnemosyne
An open-source memory engine born from Hermes. Sub-ms recall
5 followers
An open-source memory engine born from Hermes. Sub-ms recall
5 followers
Mnemosyne is a hermes-first, native, sub-millisecond memory system for AI agents using SQLite. No HTTP, no servers, no API keys. 500x faster than cloud alternatives. Open source.
Mnemosyne, the Titaness of Memory, mother of the nine Muses. Every poet and hero answered to her first.
Now she's a memory engine for your agents.
Mnemosyne started as the memory layer for Hermes, the open-source agent framework from Nous Research. Hermes agents needed somewhere to store conversations, preferences, long-running context. The existing options were either SaaS APIs or Docker+Postgres pipelines with a dozen moving parts.
So we built our own. Then open-sourced it.
What it does:
- 65.2% on BEAM at 100K scale — ICLR 2026 benchmark. We iterated from 35.4% in v2.5 to 65.2% in v3 through architecture work. Not done yet.
- Three-tier memory architecture — working memory, episodic summaries, semantic knowledge graph. All in one portable database.
- Hybrid search — vector + FTS5 + importance + temporal scoring. No separate vector DB.
- Built-in MCP server — plug into Cursor, Claude Code, Codex, Windsurf, OpenWebUI in 30 seconds.
- 23 Hermes plugin tools — lifecycle, validation, graph traversal, collaborative memory, export/import, diagnostics.
- Local-first, deployable anywhere — laptop, homelab, server.
Built for Hermes, built to last:
Mnemosyne was forged for Hermes agents by Hermes. That's where the 23-tool plugin got battle-tested. But MCP + Python SDK means Cursor, Codex, or your own stack plugs in just as easily.
The community:
800+ GitHub stars. 17K+ PyPI downloads. 26 contributors. An active Discord where people ask questions, share integrations, and push the project forward.
Memory is the last unsolved primitive in agent infrastructure. We built Mnemosyne so you don't have to rebuild it yourself!
pip install mnemosyne-memory[all]