Cortex: A Cognitive Runtime for Large Language Models Built on Cognitive Science Principles

This article introduces Cortex, an innovative cognitive runtime framework for large language models. Unlike mainstream agent frameworks, it adopts a systematic design approach based on the first principles of cognitive science, translating cognitive science theories into type-level architectural constraints (such as Global Workspace Theory, Complementary Learning Systems, etc.). Through its three-layer architecture (cognitive hardware, execution protocol, behavior library), it enables intelligent agent systems with memory, metacognition, and self-evolution capabilities, and provides a rich ecosystem of tools, interfaces, and plugins.

Current mainstream agent frameworks have made progress in areas like persistent memory and tool orchestration, but Cortex aims to elevate these capabilities from ad-hoc patchwork to systematic architectural design. Its core innovation lies in directly translating mature cognitive science theories into structural constraints enforced by the Rust compiler: Global Workspace Theory shapes the concurrency model, Complementary Learning Systems guide memory consolidation, metacognitive conflict monitoring is a first-class subsystem, Drift Diffusion Model replaces ad-hoc confidence heuristics, and Cognitive Load Theory drives context pressure responses.

Cortex is divided into three layers:

1. Bottom Layer (Cognitive Hardware)：Infrastructure encoded by the Rust type system, including event sourcing logs, a ten-state transition machine, a three-stage memory pipeline (capture → materialization → stabilization), five metacognitive detectors, a drift diffusion confidence model, three attention channels, three-level goal organization, four-axis risk assessment, etc.
2. Middle Layer (Execution Protocol)：A strategy layer that drives the bottom layer, including four prompt layers (soul layer, identity layer, behavior layer, user layer). LLM requests combine these layers and attach relevant context.
3. Top Layer (Behavior Library)：A behavior library with independent learning cycles, including five system skills (deliberate, diagnose, etc.). Skills are activated through multiple paths and their utility is tracked, with a hierarchical structure of system/instance/plugin.

Every design decision of Cortex is based on peer-reviewed theories, with the corresponding relationships as follows:

Theory	Implementation	Source
Global Workspace Theory [Baars]	Exclusive foreground rounds + log broadcasting	orchestrator.rs
Complementary Learning Systems [McClelland]	Capture → materialization → stabilization	memory/
Anterior Cingulate Conflict Monitoring [Botvinick]	Five detectors + Gratton adaptive threshold	meta/
Drift Diffusion Model [Ratcliff]	Fixed incremental evidence accumulation	confidence/
Reward Prediction Error [Schultz]	EWMA tool utility + UCB1 exploration-exploitation	meta/rpe.rs
Prefrontal Hierarchy [Koechlin]	Strategic/tactical/immediate goals	goal_store.rs
Cognitive Load Theory [Sweller]	7-region workspace + 5-level pressure	context/
Default Mode Network [Raichle]	DMN reflection + 30-minute maintenance	orchestrator.rs
ACT-R Production Rules	System/instance/plugin skills + SOAR chunking	skills/
These theories ensure that system components have a solid scientific foundation, rather than being ad-hoc heuristic methods.

Tool Categories: File I/O, execution (bash), memory operations, Web, media, delegation, scheduling, etc. These can be extended via MCP servers and native plugins. Interface Support: CLI, HTTP, JSON-RPC (multi-transport layer), instant messaging (Telegram/WhatsApp/QQ), MCP server mode, ACP mode, with Actor identity mapped across transport layers. Plugin System: A zero-dependency public API is implemented via cortex-sdk, allowing plugins to contribute tools, skills, etc. The official cortex-plugin-dev plugin turns Cortex into a coding agent, providing 32 native tools and 7 workflow skills.

Technology Stack: Rust 2024, SQLite WAL storage, Tokio asynchronous runtime, Axum HTTP framework, JSON-RPC 2.0, support for 9 LLM providers, tree-sitter parsing, libloading plugin loading. Deployment: Supports one-click script installation or source code building. Upon first launch, identity, collaborator profiles, and work protocols are established via a guided dialogue.

Cortex represents an important evolutionary direction for large language model infrastructure, shifting from feature patchwork to theory-driven systematic design. By translating cognitive science theories into architectural constraints, it provides a solid foundation for building coherent, self-correcting, goal-oriented intelligent systems. As the complexity of AI systems increases, a systematic understanding of cognitive architectures will become more important, and Cortex's theory-driven design philosophy offers a reference paradigm for future AI infrastructure development.