mnemosyne: Mnemosyne: Local-First AI Memory Layer

off_topic: false summary: Mnemosyne is a local-first, SQLite-backed memory layer for AI agents that provides working memory, episodic memory, and a temporal knowledge graph (TripleStore) without requiring external services or cloud dependencies. how_it_works: Mnemosyne uses a BEAM architecture with three tiers: working memory (hot context with TTL), episodic memory (long-term storage using sqlite-vec and FTS5 for hybrid vector and keyword search), and a TripleStore for temporal knowledge graphs. It combines vector similarity (50%), FTS5 rank (30%), and importance (20%) for scoring, and uses binary vector compression (MIB) to store embeddings efficiently in SQLite. key_features:

  • Works with multiple agent frameworks (Cursor, Claude Code, Codex, OpenWebUI, OpenClaw, Hermes, etc.) via MCP or direct SDK
  • Zero dependencies beyond Python (with optional extras for embeddings and full features)
  • Built-in MCP server for stdio or SSE transport
  • Python SDK for direct integration
  • CLI for memory operations (store, recall, stats, sleep, export/import, sync)
  • Benchmarked on LongMemEval (98.9% Recall@All@5) and BEAM (65.2% end-to-end QA at 100K scale)
  • Local-first by default with optional encrypted sync
  • Configurable via environment variables (embedding model, weights, etc.)
  • Hermes plugin for Hermes Agent users
  • Bidirectional memory sync (Mnemosyne Sync) between instances limitations:
  • BEAM benchmark results are from v3.0.0 (May 2026) and do not include later enhancements (polyphonic recall, enhanced recall, SHMR, persona tier)
  • BEAM retrieval shows a flat 20% Recall@10 across scales (low absolute value but notable for not degrading with scale)
  • LongMemEval metric for Mnemosyne (Recall@All@5) is not directly comparable to others' Recall@5
  • BEAM end-to-end QA score used a different judge (Llama 3.3 70B with DeepSeek V4 Flash) than some competitors, limiting direct comparison
  • Re-running benchmarks on the current codebase is an open task

Sources