SYNAPSE (Scalable Yet Networked Autonomous Processing and Storage Engine) is a hybrid cognitive architecture that replaces
the monolithic weight matrix of conventional large language models with a distributed network of small, independently trainable
micro-agents — called Entangled Autonomous Neurons (EAN) — operating above a frozen language model backbone.
The system integrates three original research contributions developed and tested iteratively:
1. Entangled Autonomous Neurons (EAN) — micro-agents combining a fast statistical core, bounded episodic memory
2. with trigger-based compression, and a symbolic relational reasoning graph.
Thinking Initiation Machines (TIM) — an inference-time KV-cache priming framework that constructs a warm cognitive
3. state before any query arrives.
Trigger Generation — a VQ-VAE-based memory compression system enabling lossless reconstruction from compact la-
tent codes.
A prototype was built and evaluated entirely on consumer hardware (NVIDIA RTX 5060 Ti, 16 GB VRAM), using only 2.18 GB peak.
Key Phase 2 results against a 3B frozen backbone on TriviaQA:
| Metric | Result | Target |
| Benchmark cosine similarity (EAN) | 0.9135 | ≥ Vanilla |
| Vanilla backbone similarity | 0.7868 | — |
| Domain specialisation rate | 95% (Phase 1) | >90% |
| Sequential isolation rate | 100% | 100% |
| Catastrophic forgetting rate | 0.0% | 0% |
| Real-time correction latency | 18–27 ms | <100 ms |
| Peak VRAM 2.18 GB <15 GB |
The architecture is proven to work at prototype scale. Domain specialisation [10], perfect isolation, and real-time updatability
are empirically confirmed. Conversational memory demonstrates the core concept but requires structured fact extraction to
meet the 80% target. All results are reproducible on a single consumer GPU.