# Helix AGI

A continuously running autonomous agent architecture that replaces cosine-similarity retrieval-augmented generation with a physics-simulated spatial memory, injecting far less context per turn.

Helix AGI is a multi-model agentic system built around a continuous background loop rather than a request-and-response cycle: a four-state event-driven pulse processes incoming messages, generates autonomous thought, executes tools and consolidates learning without waiting for a human prompt. In place of typical retrieval-augmented generation, it retrieves memory through a physics simulation over an eight-dimensional vector space, governed by simulated mass, distance and recency instead of cosine similarity, with no embedding API calls needed at inference time.

The project reports injecting an average of about 30 tokens of retrieved context per turn, compared with roughly 1,900 for flat semantic retrieval-augmented generation in its own comparison. It can run fully locally with Ollama models through a companion setup wizard, and its documentation includes module-by-module audits of its pulse loop, memory subsystems and an emotional-simulation component.

Website: https://github.com/munch2u-a11y/Helix-AGI