This paper proposes the Newcomen Architecture, a hybrid system combining transformer-based large language models with recurrent neural networks and vector-based memory stores to address fundamental limitations in current LLM implementations; specifically the absence of persistent conversational state, episodic memory, and emotional continuity. The architecture maps to neuroscientific structures: the transformer as cortex (interpretation), the RNN as hypothalamus/thalamus (state and salience), and vector databases as hippocampus (memory). We present a complete implementation blueprint using off-the-shelf components and note convergent independent work from Google Research's Titans architecture, which arrives at a similar tri-component decomposition through mathematical formalism. The Newcomen Architecture offers a practical retrofit path for augmenting existing deployed LLMs with persistent state and memory.
Building similarity graph...
Analyzing shared references across papers
Loading...
Justin Dew
Building similarity graph...
Analyzing shared references across papers
Loading...
Justin Dew (Sat,) studied this question.
www.synapsesocial.com/papers/699d4008de8e28729cf6503c — DOI: https://doi.org/10.5281/zenodo.18728294