Unmetered memory as the missing layer.
What if we've been looking for memory in the wrong place?
Most artificial neural networks learn by changing the strength of connections between neurons. Once training ends, those connections typically remain fixed. The artificial neuron processes information, produces an output, and retains no persistent record of its own experience.
For decades, neuroscience has treated synaptic connections as central to how memories are stored. Yet a growing body of research suggests that the story is more complicated. Memories can survive significant disruption of the synapses associated with learning. And in 2018, David Glanzman's laboratory at UCLA demonstrated something extraordinary: RNA extracted from trained sea slugs could induce a learned defensive response in animals that had never undergone the training themselves.
The findings challenged the idea that memory resides exclusively in the connections between neurons. They also raised a question with profound implications for artificial intelligence: what if the neuron itself could remember?
NURESA is a computational architecture built around that question.
Rather than treating artificial neurons as stateless processors, NURESA gives each neuron its own persistent memory: a register that accumulates experience and a vault that consolidates what the neuron has learned.
That single architectural change unlocks everything else.
Because learned capability can be consolidated into a distinct memory state, it can be extracted. Because it can be extracted, it can be packaged and transferred. A compact memory packet can be installed in another model of the same architecture, enabling it to acquire learned capabilities without additional training, access to the original training data, or gradient updates.
In Neurons Are All You Need, we demonstrate this radical idea in a proof-of-concept set of experiments across three model families: a multilayer perceptron, GPT-2, and Qwen2.5 1.5B. The memory system autonomously identifies the most important neurons, and the transferred packets protect against catastrophic forgetting while preserving training data privacy. This interactive essay explains the biological inspiration, the architecture, the reported results, and the unanswered questions.
A conventional feedforward artificial neuron does not maintain an autonomous persistent memory register. It receives inputs, multiplies by weights, applies a function, outputs a value. When training stops, the neuron retains no record of what it did. Durable learned parameters are primarily represented in weights. Previous attempts to add memory to neural networks (Neural Turing Machines, memory-augmented networks) attached an external memory bank, a centralized scratchpad, whether text or tokens, that the network reads and writes through attention. It's still outside the neuron.
NURESA takes the opposite approach. Every single neuron gets two additions: a persistent register M that accumulates a running history of the neuron's activity across training, and a vault C that consolidates the neuron's learned configuration once M crosses a commitment boundary. No external controller. No centralized memory. Decentralized, persistent, and transferable.
M doesn't just track activity. It changes how the neuron behaves. As M accumulates, each neuron blends between two response modes: sharp and reactive (like a standard ReLU) at low M, smoother and bounded (like a sigmoid) at high M. This gives the network temporal dynamics without needing recurrent connections. M also gates plasticity: committed neurons have their learning rate substantially reduced, creating two populations during training. Frozen experts and still-learning generalists.
When enough M accumulates, the neuron's weights are consolidated into C via a smoothing window. The vault is a separate, stable copy of what that neuron learned. It survives erasure of the original weights. And it can be extracted and moved.
Train Model A. The most active neurons commit and consolidate. Extract their vaults into a compact memory packet. Inject the packet into a fresh Model B of the same architecture. Model B gains the capability immediately. No training, no data, no gradient updates.
NURESA was tested on a simple classifier (MLP, 269K parameters), a mid-range language model (GPT-2, 345M), and a modern LLM (Qwen2.5, 1.5B). In every case, a memory packet extracted from a trained model and injected into a fresh one produced immediate capability gains with zero training.
Vault injected into a randomly initialized network. No training. Random and shuffled packets degraded to chance, confirming the transferred information has neuron-specific structure.
345M parameters. Perplexity dropped from 254 to 57 with vault transfer. Scrambled packets made it worse.
At 50% commitment, GSM8K zero-shot accuracy reached 55.8%, approximately 88.15% of the supervised fine-tuning reference. The recipient received no additional gradient updates.
At 30% commitment, M-selected neurons recovered 91.5% of full fine-tuning performance on GSM8K. Random selections and largest-weight-change selections at the same 30% level recovered at most 28.5%. The memory register isn't a bookkeeping device. It's autonomously identifying the computational core of what the model learned: a sparse population of units that captures disproportionate capability.
The approximately 30% selection point is an interesting parallel with sparse biological engram populations discussed in the paper, but the correspondence is suggestive rather than evidence of a shared mechanism.
Across the experiments, M+C packets were smaller than their corresponding full models. For Qwen2.5 1.5B, the paper reports up to 32.7× size reduction and transfer times of 65.1–78.0 seconds, with up to a 48.9× speed-up relative to the reported fine-tuning comparison. For the MLP, transfer was 580,000–800,000× faster than its fine-tuning comparison. These timings depend on the specific experimental setup (Figures 8–10).
These are distinct experimental comparisons, not one common commitment setting. See Figures 4 and 5 in the arXiv paper for full baselines and controls.
Forgetting protection. Catastrophic forgetting, one of the oldest problems in deep learning, happens because fine-tuning overwrites old weights to learn new things. In the tested conditions, NURESA mitigated this form of forgetting. The recipient receives an M+C overlay without additional gradient updates; the tested benchmarks showed less domain loss than fine-tuning. In Qwen2.5, fine-tuning on math caused a 9-point drop on MATH-500; vault transfer preserved baseline scores. Across six other Qwen2.5 benchmarks, the paper reports no significant negative effects.
Privacy. For Qwen2.5, the tested membership-inference attack yielded AUROC 0.4769 for M+C versus 0.8677 with full-weight access (Figure 12). These results are promising for privacy, but do not establish immunity to other attacks.
Full experimental details: arXiv 2611.00186.
The road ahead · Our perspective
When Tesla shipped its first Roadster in 2008, the car worked. What didn't exist was the infrastructure. No charging network, no service centers, no supply chain for battery packs. The product was ahead of the ecosystem. Adoption didn't reach critical mass until the infrastructure caught up: Supercharger stations, battery gigafactories, regulatory frameworks for EVs.
AI infrastructure faces an analogous problem. Much of today’s AI tooling is optimized around weight-based computation rather than autonomous persistent memory inside each computational unit. PyTorch's computational graph, NVIDIA's tensor cores, ONNX export formats, model serving frameworks, quantization pipelines. All of it optimized for matrix multiplication on weight matrices. Native support for the particular per-unit persistent-state memory mechanisms, free from the cost of memory tokens or scaffold prompts, are not the standard in mainstream deployment stacks.
Today, NURESA works by retrofitting onto existing architectures. M and C are bolted onto standard layers. These experiments establish a working path. The next opportunity is to design around persistent neuron state from the start. A native implementation, where stateful neurons are first-class primitives in the framework, the compiler, and eventually the hardware, would look very different, and would unlock capabilities that retrofitting can't reach.
This is both the limitation and the opportunity. The infrastructure gap is real, and it means there are models and deployment scenarios where NURESA won't play well today. But it also means that the entire AI stack, from frameworks to silicon, was designed around an assumption that may be incomplete. If persistent intra-unit memory proves useful at scale, parts of the stack may need to evolve.
The road ahead · Our perspective
Energy, chips, infrastructure, models, applications. Trillions of dollars of value organized into layers. Every layer is about building, training, or deploying models. None of them address what happens after a model learns.
Our hypothesis is that persistent memory could become a distinct capability layer between models and applications. NURESA offers an early experiment in packaging and transferring learned state between compatible models. Composition, general-purpose interoperability, and comprehensive privacy protection remain open research questions.
M selects neurons during training, not after. Does that produce better vaults than simply picking units with the largest weight changes?
Current experiments use recipients that share the same pretrained base. What happens with a different-seed recipient?
Results go to 1.5B parameters. What happens at 4B? At 70B? Scaling these mechanisms is a central part of the research roadmap.
Math from one training run, code from another, medical reasoning from a third, all injected into the same base model. Does it work? Do they interfere?
NURESA today retrofits onto transformers built for stateless neurons. What does a network look like when every unit is designed from the start to accumulate, consolidate, and share memory? New frameworks. New training loops. New memory-capable agent swarms. Potentially new hardware abstractions.
This is an active research program. We will publish new experiments, replication attempts, negative results, and architectural advances as the evidence develops. The research record begins with the foundational paper.
The initial paper introduces NURESA and reports experiments across MLP, GPT-2, and Qwen2.5 1.5B, including transfer, commitment selection, forgetting, compression, and privacy evaluations.
Read the primary paperTransfer between compatible models; performance improvements without recipient retraining; benchmark-specific forgetting mitigation; promising results under two tested privacy attacks.
Cross-architecture transfer, transfer between arbitrary initializations, robust skill composition, production-scale continual learning, and general privacy guarantees.
Research collaborations, replication inquiries, and technical discussion are welcome. For code and data access, the paper identifies UCLA Technology Development Group, as the contact under the current IP arrangements. You can also reach the research team at research@nuresa.ai.
Read the paperContact the research team
An independent interactive explanation of the research. Interpretations and future-looking ideas are distinguished from reported results.
Primary paper · UCLA TDG · Research correspondence
© 2026 Joseph B. Alzagatiti, Alain C. Glanzman, Perry A. Demsko, Pablo I. Salanova & David L. Glanzman