Mnemosyne
Auditable, local-first memory for AI agents in one SQLite file. MCP server for Claude Code, Cursor and Codex. npm: mnemosy-ai
Mnemosyne is a local-first, long-term memory for AI agents that stores learned information with evidence in a single SQLite file and serves it over MCP, SDK, CLI, and HTTP. The software is MIT licensed with no software license fee. It achieved a measured accuracy of 409 out of 500 questions on the LongMemEval-S benchmark.
Version 3.0 adds outcome-gated consolidation, dependency inference for review, a poisoning gate, opt-in precision controls, and hybrid-path bisect. It can be used via TypeScript, Python HTTP client, CLI, MCP, or authenticated HTTP.
Summary drafted from the project's own website. Every sentence is backed by text on that page and was reviewed before publishing.