Drift: A Biologically-Grounded Cognitive Architecture for Persistent LLM Cognition

Large language models are stateless by design — each session begins as a blank slate, with memory provided only through context windows or external retrieval. We present Drift , a cognitive architecture that endows stateless LLMs with persistent, biologically-grounded cognition across sessions. The system comprises 112 Python modules (60,000 lines) implementing a 19-stage retrieval pipeline, affect-modulated search, reinforcement-learned pipeline optimization, and cryptographically attested identity evolution. Developed as a living system — with two independent agents operating autonomously across 8+ platforms and environments over 200+ sessions — the architecture has been validated through continuous real-world use rather than synthetic benchmarks alone. We identify six genuinely novel contributions: (1) per-stage Q-learning that treats retrieval pipeline optimization as a multi-armed bandit problem, (2) per-memory Q-learning that treats individual memories as bandit arms whose utility is learned through retrieval feedback, (3) topology-based cognitive fingerprinting that derives identity from co-occurrence graph structure rather than memory content, (4) rejection logs as a cryptographically attestable identity signal, (5) predictive coding applied to memory retrieval using Rescorla–Wagner learning, and (6) spring-damper mood dynamics where velocity represents felt emotion. The system has been running in production with verifiable session-to-session continuity. An independent 10-agent specialist review scored the architecture 7.0/10 for theoretical coherence and HIGH for novelty, identifying it as field-leading in affect integration and identity persistence. We present the architecture, discuss its biological grounding, acknowledge its limitations, and propose a rigorous testing protocol. This paper serves as both a systems description and a call for empirical validation of biologically-grounded LLM cognition.