hermes-labs-ai/zer0dex

zer0dex is a local memory system for AI agents that need to remember information across many conversations or sessions. The problem it addresses is that AI agents typically start each session with a blank slate, or they store everything in one giant text file and still struggle to find the right information at the right moment, or they rely on a vector search that finds isolated facts without understanding how those facts relate to each other. The system works in two layers. The first is a compressed, human readable markdown index file that stays loaded in the agent's context at all times, acting like a table of contents that tells the agent what categories of knowledge exist without storing every detail. The second layer is a local vector store, built on mem0 and chromadb, that holds extracted facts from logs, conversations, and notes, and is searched by meaning rather than exact keywords. A lightweight local server keeps this vector store warm and automatically runs a query before every message the agent receives, injecting the top matching memories into its context without anyone having to decide manually what is relevant. According to the project's own benchmark of 97 test cases, this two layer approach reached 91.2% recall, compared to 52.2% for a flat memory file and 80.3% for vector search alone. The gap was largest on cross reference questions, ones that require connecting facts from two different topics, where the combined approach scored 80% against 37.5% for vector search by itself. The whole system runs locally with no ongoing hosting cost. You would use this if you are building or running a persistent AI agent that needs to remember past conversations, project details, or user preferences over long periods, without paying for a hosted memory service. It is written in Python, installed with pip, and requires Ollama, a tool for running AI models locally, along with an embedding model and an extraction model pulled through it.