Lethe drops persistent DuckDB memory for Claude Code

Lethe solves the agent amnesia problem by moving beyond simple chat logs to a high-performance, local-first analytical memory layer.

• –DuckDB per-project storage enables zero-dependency indexing and cross-project querying via simple SQL ATTACH commands.
• –Hybrid pipeline (BM25 + dense + cross-encoder) ensures high-precision retrieval of technical details, outperforming standard vector-only RAG.
• –Performance gains of +0.017 NDCG@10 on LongMemEval are driven by a learned suppression layer tuned for long-term conversational memory.
• –Research indicates the underlying retrieval mechanism is highly specialized, as benchmarks on the NFCorpus dataset showed the performance gains did not transfer to general IR tasks.
• –MCP-ready architecture allows Claude Code to automatically write session summaries and recall past decisions without manual context management.