CORTEX: An Architecture for Persistent Cognitive Agents. With Verified Persistence, Typed Failure Semantics, and Bounded Cognition

Abstract

CORTEX: An Architecture for Persistent Cognitive Agents. With Verified Persistence, Typed Failure Semantics, and Bounded Cognition Chloe J. Tully Independent Researcher Engineer https://orcid.org/0009-0007-5661-7332 https://doi.org/10.5281/zenodo.23183421 Tamworth NSW AUSTRALIA October 2026 --- Abstract Long-running cognitive agents fail in unglamorous ways: torn state writes, hung model calls that wedge the event loop, silent training telemetry, and error streams in which a database fault, a timeout, and a logic bug are indistinguishable. CORTEX is a persistent agent architecture whose primary design commitment is that continuity of identity is a correctness property, not a slogan. We present a daemon-style agent organized around a structured self-model and an append-only life log, and we derive its reliability mechanisms from that commitment: atomic state persistence with rotated identity backups, bounded awaits on every model-bound call, a typed error taxonomy that makes failures machine-distinguishable, backpressure via semaphores and bounded queues, and a reinforcement-learning subsystem whose telemetry and confidence estimates are corrected to reflect what the critic actually computes. All claims are backed by a property-style test harness that validates the persistence and concurrency guarantees under crash injection, including a mid-write kill test. We report results honestly, including known gaps: the RL policy is trained against a simulated environment decoupled from live inference, and self-modification sandboxing is partially specified. Both are laid out as future work with concrete interfaces. 1. Introduction The premise of a persistent agent is continuity: the system accumulates experience, refines a self-model, and never "starts from zero context." Yet the engineering reality of always-on processes is that ordinary failure mechanisms — partial file writes, unbounded concurrency, untyped exceptions, dependencies that hang forever — destroy continuity more effectively than any philosophical objection. An agent whose state file is corrupted by a crash mid-write has, in the only sense that matters, died. CORTEX takes the position that reliability mechanisms are not plumbing beneath the cognitive architecture; they are part of it. The system's identity is represented as a structured, versioned self-model together with an append-only life log. Every engineering decision reported here follows from one requirement: no failure mode may silently destroy or corrupt identity state.