Finch
Project
2025 - 2026Finch is a vector database written in Rust that runs inside the calling process, plus a memory layer for agents. The database stores documents with dense or sparse vectors on disk. The memory layer stores episodes and claims in Postgres with pgvector, tracks when each claim held, and packs current state with the source text that proves it.
It scores 71.0% on the MEME benchmark under the
paper's scoring (493 / 694; 74.6% as judged by the official evaluator) and 73.43
combined on EnterpriseRAG-Bench. MEME extraction
and selection used gpt-6-luna, and answers used gpt-4.1-mini under the official
protocol. On EnterpriseRAG-Bench, gpt-5.6-luna expanded queries and wrote answers,
and Cohere Rerank 3.5 reranked results.
Other systems come from the EnterpriseRAG-Bench leaderboard and the MEME paper. MEME scores use the paper's scoring.
Finch per-category MEME scores.
Currently in the process of writing a detailed breakdown which will be called Memory for AI where I conduct a mini case study covering current memory systems, inspiration for the project, how I chose to measure success/performance, latency optimizations, effects of the LLM you choose to use in this memory system, and more.