Lybrary brings persistent code memory to AI agents via local MCP server

AI coding agents face a persistent inefficiency: every time they start a session, they re-read the same files to understand the codebase structure, burning tokens and filling context before writing a single line. On large codebases, this pattern repeats dozens of times per session. Lybrary addresses this by building a persistent, queryable memory indexed along actual code boundaries—functions, classes, methods—rather than arbitrary line splits. The tool uses tree-sitter, an incremental parser, to respect AST structure when chunking, ensuring that a semantic unit like a function stays intact.

The architecture is deliberately local-first: embeddings run via ONNX Runtime using MiniLM-L6-v2, SQLite stores the chunks and vectors, and a background daemon watches for file changes and re-indexes only what changed. By reducing multi-file reads to a single semantic search query, agents cut token waste by 80–90% on large codebases. Integration as an MCP server means agents in Cursor, Claude Desktop, Windsurf, and Kiro can call `memory_query` natively before attempting file reads, treating the persistent memory as a first-class tool rather than an afterthought.