I have developed a learning-association thinking system based on simulations of synaptic connections in the human brain. I am currently refining it to truly mimic the way the human brain thinks. It aims to form genuine conceptual understanding of words and perform real reasoning, rather than token prediction as done by LLMs. The project is still under development. The first minimal prototype is available at: https://github.com/zhj-ldm/CogniBrain
CogniBrain is a pure bionic human cognitive underlying system that completely abandons the traditional Transformer LLM probabilistic generation paradigm. Unlike mainstream AI that generates text through statistical word prediction, this system simulates real human brain neural logic, relying on neuron nodes and synaptic connection plasticity to implement autonomous cognition, associative learning, and step-by-step reasoning.
Important clarification for the current version: This iteration is a** pure cognitive thinking core**. It only completes internal thinking, reasoning, memory and incremental learning. It does not support complete fluent active dialogue yet. This is a dedicated underlying cognitive module, separated from the advanced human-computer interaction layer.
A brand-new, advanced full conversational version is currently under independent development. The upgraded version will be docked to this original bionic cognitive core, supporting real human-like dialogue, complex interactive reasoning and multi-scene intelligent interaction, forming a complete closed-loop bionic AI system.
The system builds independent conceptual neuron nodes for all input information, including entities, attributes, logical relationships, causality and scene rules. All nodes grow, split and iterate dynamically with learning content, without fixed preset structures.
Bidirectional activation links are established between nodes through bionic synaptic connections. Synaptic weights are adjusted in real time according to activation frequency and association times, strictly following Hebbian learning: frequently co-activated connections are strengthened, while long-term inactive connections decay gradually.
Different from LLM’s one-pass synchronous generation, CogniBrain adopts human-like decomposed thinking logic. It splits complex problems into multiple small reasoning units, stores temporary results through working memory, and completes progressive logical derivation step by step.
During the thinking process, the system automatically activates related concepts, detects and eliminates logical conflicts, and finally converges a complete, self-consistent cognitive conclusion. There is no random splicing, no speculative generation, and zero cognitive hallucinations.
The system initializes with a blank cognitive state, with no built-in pre-trained data or fixed knowledge. All cognition comes from autonomous parsing and learning of external input information.
It extracts concepts, relationships and logical rules from input content, continuously optimizes and supplements the internal neural network. New knowledge iterates and strengthens the original cognition instead of overwriting it, supporting analogy generalization and scenario migration judgment.
Simulate real human memory attenuation rules. Frequently activated core concepts and synaptic connections are solidified into long-term stable memory. Infrequently used and low-correlation connections decay slowly and are forgotten automatically, avoiding the rigid memory accumulation problem of traditional AI models.
Traditional LLM: Relies on massive pre-training corpora, generates content by probability token prediction, performs pseudo-reasoning without real conceptual understanding, and suffers from serious logical hallucinations and factual errors.
CogniBrain Cognitive Core (Current Version): Zero pre-training, real neural association thinking, rigorous step-by-step reasoning, autonomous incremental learning, no hallucinations, completely restoring human cognitive logic.
The current open-source version is positioned as the underlying pure cognitive foundation, responsible for all core thinking, reasoning, learning and memory capabilities, equipped with complete GUI visualization, persistent brain state storage and interactive training functions.
The upcoming official full conversational version is in active development. It will build an independent human-computer dialogue interaction layer based on this cognitive core, realizing real thinking-based dialogue (not text splicing), supporting complex multi-turn interaction, active communication and advanced cognitive reasoning.