codenamekt/hexus

Hexus is a memory system for AI agents that stores and retrieves information using your own PostgreSQL database instead of a paid cloud service. It converts text into vector embeddings using a small local language model (MiniLM-L6-v2), stores them in Postgres with the pgvector extension, and searches them at millisecond speed when an agent needs to remember something. Because all embedding happens locally on your CPU, there are no calls to an external AI service at search time. The project serves two kinds of users. If you are running a Hermes agent (an open-source multi-agent framework), you install Hexus as a Python plugin and it automatically gives each agent its own scoped memory space, called a theme, so separate agents do not overwrite each other's notes. If you want to give Claude Desktop, Cursor, or another MCP-compatible AI tool persistent memory, you run Hexus as a standalone Model Context Protocol server, either via Docker or by pointing Claude's config file at the Docker image. Beyond basic vector search, the tool includes hybrid search that combines vector similarity with keyword matching (BM25), a temporal decay feature that makes older memories less prominent over time, a TTL option that automatically forgets entries after a set period, entity tagging, and conversation summaries. The README reports a single embed taking 7.4 milliseconds and a top-5 recall taking 2 milliseconds on a basic CPU with no GPU. Setup requires a running PostgreSQL 16 instance with the pgvector extension enabled. A Docker Compose file is included to spin one up quickly. The Hermes plugin path is a single pip install. The MCP server path involves adding a short JSON block to Claude's desktop config file. Hexus is licensed under the BSD 3-Clause license and described as intentionally minimal: a drop-in storage and search layer you add to an existing agent setup rather than a complete agent platform.