Mnemosyne

Mnemosyne is an open source, local first memory system for AI agents. It gives an agent persistent memory stored in SQLite, installed with a single pip command and requiring one pure Python dependency, with no servers, Docker, cloud account or API keys. Queries are described as running in under a millisecond, and all data stays on the user's machine. Developers call simple remember and recall functions from Python to store memories with an importance score and scope and to retrieve relevant context later. Search combines vector similarity using sqlite vec, full text search and importance scoring in a hybrid ranking. Memory is organized in three tiers: working memory for hot context, episodic memory for long term storage, and a scratchpad for reasoning. Older working memories are summarized and moved into episodic storage automatically through configurable sleep cycles, and ignore patterns keep noisy content out. The site reports benchmark results on the BEAM memory benchmark and offers a browser demo that simulates the behavior. It is intended for developers building agents and works with Hermes Agent, Claude Code, Cursor, Codex, OpenWebUI and OpenClaw, including through MCP. The project is free and listed as free forever. It is run as an Open Source Collective on OpenCollective, where supporters can contribute.