# YantrikDB

A cognitive memory database for AI agents that actively manages what it stores, consolidating duplicates, detecting contradictions, and fading out stale memories through temporal decay.

YantrikDB is described as a cognitive memory database for AI agents, addressing a specific limitation of plain vector databases: they store memories but do not manage them, so recall quality degrades once a collection grows into the thousands of entries with no consolidation, forgetting, or conflict resolution. It can be embedded as a library, run as a standalone server, or reached over MCP from clients such as Claude Code, Cursor, or Windsurf.

Individual memories carry an importance weight and a half-life, so their relevance score decays over time and eventually stops surfacing in recall unless a query explicitly asks for expired memories. The project reports substantially lower per-query token cost than keeping the same memories in a flat Markdown file, with the gap widening as the memory count grows, based on a reproducible benchmark script in a companion repository. It is written in Rust, Apache 2.0 licensed, and can run clustered over HTTP.

Website: https://yantrikdb.com