Agent long-term memory that fits in one SQLite file

What it does

Mnemosyne is a persistent memory layer for LLM agents such as Cursor, Claude Code, and OpenWebUI. It stores facts and conversation context in a single SQLite file using a three-tier architecture: hot working memory with TTL eviction, long-term episodic storage, and a versioned temporal knowledge graph called TripleStore. Agents interact with it through a built-in MCP server, a Python SDK, or native plugins, and retrieval uses a hybrid of vector similarity, FTS5 full-text rank, and importance scoring.

The interesting bit

Rather than outsourcing vectors to a dedicated database, Mnemosyne compresses 384-dimensional float32 embeddings into 48-byte binary vectors using an information-theoretic binarization scheme called MIB, then performs Hamming-distance search entirely inside SQLite. The README claims this yields a 32× size reduction and 9.4× episodic storage savings, though it also reports that recall@10 stays flat at 20% across all tested scales.

Key highlights

• Built-in MCP server and integration templates for eight-plus platforms, including Windsurf, OpenClaw, and OpenWebUI.
• Scores 98.9% Recall@All@5 on LongMemEval and 65.2% on BEAM’s 100K end-to-end QA benchmark.
• Hybrid retrieval scoring is split 50% vector similarity, 30% FTS5 rank, and 20% importance, all computed within SQLite.
• Claims 100% abstention accuracy, meaning it reportedly returns no answer rather than hallucinating when context is missing.
• Working memory auto-injects before LLM calls, while episodic compression consolidates older facts into summaries.

Caveats

• The project bills itself as “zero-dependency” and requiring “no external services,” yet the environment configuration table hardcodes defaults for an external embedding API endpoint and key, implying outbound calls are expected.
• The “sub-millisecond” branding clashes with the benchmark table, which lists query latencies between 372 ms and 493 ms for smaller datasets and an anomalous 35 ms at the 10-million-message scale.
• On the BEAM end-to-end benchmark at 100K scale, Mnemosyne trails Hindsight (73.4%) despite beating Honcho and LIGHT.

Verdict

Agent builders who want a SQLite-native, MCP-ready memory system with published benchmark scores should try this. If you need a fully air-gapped setup with no embedding API dependencies, or if you require the absolute highest BEAM QA scores, the README’s own numbers suggest looking elsewhere.