trvon/yams

Persistent memory for LLMs and apps. Content-addressed storage with dedupe, compression, full-text and vector search.

What it solves

YAMS provides a persistent memory system for LLMs and applications, solving the problem of efficiently storing, deduplicating, and retrieving large amounts of data for AI assistants. It combines content-addressed storage with both traditional keyword search and modern semantic search to ensure data is stored once and retrieved accurately.

How it works

The system uses SHA-256 content-addressed storage with Rabin chunking for block-level deduplication and compression. It integrates SQLite FTS5 for full-text search and embeddings for vector search. To handle structured data, it uses Tree-sitter for symbol extraction across 18 different programming languages. It also features a Model Context Protocol (MCP) server, allowing AI assistants to interact with the memory system via JSON-RPC.

Who it’s for

Developers building AI agents and LLM-powered applications who need a high-performance, durable memory layer that supports semantic retrieval and efficient storage of code and documents.

Highlights

  • Hybrid Search: Combines full-text search and semantic vector search for flexible retrieval.
  • Efficient Storage: Implements block-level deduplication and compression to minimize disk usage.
  • MCP Integration: Includes a built-in MCP server for seamless connection to AI assistants.
  • Hardware Acceleration: Supports GPU acceleration via ONNX for various platforms including NVIDIA, AMD, and Apple Silicon.
  • Code Awareness: Extracts symbols from 18 languages using Tree-sitter.