Mnemosyne is a native memory system for AI agents. Persistent, structured, and context-aware — built on SQLite with vector search, full-text retrieval, and biological-inspired BEAM architecture.
Biological-inspired memory tiers: Working, Episodic, Semantic, and Scratchpad — each optimized for different access patterns.
Combine dense vector similarity with SQLite FTS5 full-text search for results that are both relevant and precise.
Zero external dependencies. Single-file database with ACID transactions, portability, and proven reliability.
All data stays local. Optional encryption at rest. No cloud required, no data exfiltration.
Optimized indexing and caching deliver median retrieval latency under 100ms for typical workloads.
First-class plugin for the Hermes agent framework. Auto-context injection, tool schema, and seamless operation.
Start with the Quick Start guide and have Mnemosyne running in under 5 minutes.
Quick Start