MIRIX

Why we included this project

Most agent frameworks treat memory as an afterthought, so assistants forget what happened last session and repeat themselves; MIRIX is built around that problem, watching what happens on screen and filing what it sees into a structured memory your agent can search later. Stored information is split into six memory types: personal facts, episodic activities, semantic knowledge, procedures, resources, and credentials, and dedicated sub-agents handle writing to and querying those stores. Retrieval pairs PostgreSQL's native full-text search with embedding-based vector similarity, so a query finds both exact keyword matches and related semantic content. The backend self-hosts through a Docker quick-start, and a Python client lets you attach the memory to agents you already run instead of committing to a proprietary assistant, while long-term data stays local under user-controlled privacy settings.