Mnemosyne

Mnemosyne is a native AI memory system for Hermes Agent and other AI agents. It is described as a universal memory layer for AI agents that need fast, reliable, persistent memory, and it is built around SQLite with a zero-dependency design.

The system emphasizes local storage and low-latency retrieval. It uses direct SQLite access for sub-millisecond queries, and the page states that all data stays on the user’s machine or in the browser in the demo. Mnemosyne also presents itself as 100% local and private, with no cloud services, no external services, and no network roundtrips. A browser demo is available, but it is explicitly a client-side simulation rather than the real install.

Its named features include native vector search through sqlite-vec, hybrid ranking that combines vector similarity, full-text search, and importance scoring, and BEAM, a three-tier memory architecture with working memory for hot context, episodic memory for long-term storage, and a scratchpad for reasoning. It also supports automatic consolidation through sleep cycles, real-time incremental updates via DeltaSync, streaming results as they arrive, and ignore_patterns for filtering noisy or irrelevant content. The page says it works offline and includes examples of using remember and recall, and it also lists compatibility with Hermes, Claude Code, Cursor, Codex, OpenWebUI, and OpenClaw + MCP.

Mnemosyne is delivered with one pip install, and the install command shown is mnemosyne-memory. The page says there are no configuration files, no environment variables, and no cloud accounts required. It is free and open source, and it is now an OpenSource Collective on OpenCollective. The bottom-line comparison section describes the tool as intended for individual developers and local agents, and it notes that users manage their own backup as a single SQLite file.