Project null-drift
A cognitive memory system for AI agents that projects semantic embeddings into a high-dimensional phase space to give agents persistent, continuous episodic memory at constant cost, instead of growing linearly with context.
null-drift projects standard semantic embeddings into a 10,000-dimensional bipolar phase space and binds sequences of events into a continuous temporal energy landscape, using fractional salience superposition and temporal permutation to build causal chains while filtering out low-salience noise. The aim is to give agents persistent memory at a flat computational cost, instead of the growing token and billing cost of accumulating raw context.
The system has three parts: a Rust daemon that manages many isolated per-thread memory environments in bounded memory and adapts to different embedding dimensions; a Rust microservice that generates embeddings locally with fastembed over ONNX Runtime and forwards them to the daemon; and SDK adapters for CrewAI and LangGraph that let agents use the memory store, including with a user's own or a locally hosted embedding model.
It is deployed with Docker Compose, exposing a multi-tenant HTTP gateway with per-thread isolation.