EMBER: Autonomous Cognitive Behaviour from Learned Spiking Neural Network Dynamics in a Hybrid LLM Architecture

Researcher William Savage has introduced EMBER (Experience-Modulated Biologically-inspired Emergent Reasoning), a novel hybrid AI architecture that fundamentally rethinks the relationship between large language models (LLMs) and memory. Instead of using an LLM as the central processor augmented with retrieval tools, EMBER places the LLM as a replaceable reasoning engine within a persistent, biologically-inspired associative substrate. This substrate is a 220,000-neuron spiking neural network (SNN) featuring spike-timing-dependent plasticity (STDP), a four-layer hierarchical organization, and reward-modulated learning. A novel encoding method allows text embeddings to be fed into the SNN, achieving 82.2% discrimination retention across different embedding sizes.

The core innovation is the SNN's ability to drive autonomous behavior. Through STDP and lateral propagation during idle operation, the network can trigger and shape LLM actions without any external prompting or scripted triggers. The SNN determines *when* to act and *what* associations to surface, while the LLM handles *how* to act by selecting an action type and generating content. In a striking demonstration, the system autonomously initiated contact with a user after learned person-topic associations fired laterally during an 8-hour idle period. Remarkably, from a clean start with zero learned weights, the first SNN-triggered action occurred after only 7 conversational exchanges (14 messages). This represents a significant step towards AI systems with persistent, evolving internal states that can drive proactive behavior.