A Portable AI Memory Layer Transfers Some Learned Skills Without Retraining

Could an AI model pass along a newly learned skill without sending its training examples or an entire updated model? Researchers reporting work performed at UCLA describe a memory-layer approach that transfers selected learned states into a matching recipient model, improving its performance without further training.

Neural networks typically acquire capabilities by changing numerical parameters called weights. In the preprint Neurons Are All You Need, Joseph B. Alzagatiti and colleagues propose an additional place to organize learned information: persistent memory associated with individual computational units, or artificial neurons.

Their system, NURESA—Neural Unit with Retentive Self-Adaptive Architecture—collects selected learned states into a transferable package. The authors report that applying this package improves a recipient model without additional gradient updates or access to the source training examples. However, the transfer currently requires source and recipient models with the same architecture and size.

What the memory package contains

NURESA gives each unit a register, called M, that accumulates a numerical record of its activation history. That record helps control the unit’s response and local adaptation. When accumulated activity crosses a commitment threshold, the unit’s learned configuration is consolidated into a separate store, called C.

Together, these stores form an M+C packet. Rather than appending notes to a chatbot’s prompt, the method applies stored states within the network itself.

Importantly, this is not learning without any weight-based training. The source models undergo conventional training or fine-tuning, and consolidation includes learned weight parameters. The proposed distinction is how selected states are accumulated, stored and transferred—not the elimination of numerical parameters or the original training process.

The authors draw inspiration from biological research they interpret as evidence for memory mechanisms beyond synaptic connections. That interpretation motivates the architecture; the AI experiments do not themselves establish how biological memory is stored.

Improvements, but incomplete transfer

The researchers tested three settings: a small network recognizing handwritten digits, GPT-2 Medium, and the 1.5-billion-parameter Qwen2.5 language model. They varied the proportion of units committed to the transfer packet from 10% to 50%.

For handwritten-digit recognition on MNIST, an untrained recipient started at approximately 11% accuracy. Transferring the packet at 50% commitment raised accuracy to 42%. That is a substantial improvement, but far below the source network’s 98% accuracy.

GPT-2 showed a similar pattern on held-out WikiText passages. Its perplexity—a measure of predictive uncertainty, where lower is better—fell from 253.85 for the pretrained recipient to 56.83 after the largest tested transfer. The fine-tuned source achieved 11.58. Here, the recipient was already pretrained: “without retraining” does not mean a language model learned from scratch.

For Qwen2.5, the team transferred states accumulated during mathematics fine-tuning. Across the commitment sweep, accuracy on GSM8K arithmetic word problems increased from 47.31% at 10% commitment to 55.8% at 50%. The authors describe the latter as 88.15% of the fine-tuned model’s benchmark score—not 88.15% of its improvement over the base model.

A separate selection experiment favored units chosen using accumulated memory over random selections or units with the largest activation signals. The memory-selected packet reached approximately 91.5% of the fine-tuned source’s score in that experiment.

Speed and retained abilities

Qwen2.5 packets ranged from about 90 to 452 megabytes, compared with roughly 2,955 megabytes for the base model. Applying them took 65–78 seconds, versus 3,187 seconds for the reported full fine-tuning run.

These timings compare applying an already-created packet with fresh fine-tuning. They do not remove the cost of training the source, and the recipient still needs its underlying base model. The packet is an overlay, not a standalone replacement.

The authors also report less forgetting on selected evaluations. Mathematics fine-tuning reduced Qwen2.5’s MATH-500 performance by roughly nine percentage points, while packet-equipped recipients retained base-model performance. GPT-2 showed a similar advantage on LAMBADA, a word-prediction test. These findings concern particular benchmarks, not proof that the method prevents forgetting generally.

Privacy and practical limits

Privacy tests were encouraging but narrow. A representation-reconstruction attack performed similarly to a random-vector comparison. A membership classifier, attempting to identify training examples, scored 0.4769 on Qwen2.5 packets using a metric where 0.5 represents chance. GPT-2 packets scored 0.5812, compared with 0.8107 for full-model access.

The authors acknowledge that these two attack types cannot rule out other leakage. They also identify commitment-threshold tuning as critical: committing too early can undermine the specificity of stored signals.

Implementation details, code and trained weights are currently subject to an access-request process administered through UCLA’s technology-development office because of a patent application. Two authors are named as co-inventors.

The preprint therefore offers evidence for selective, portable transfer of learned states under tested conditions—not unlimited memory, universal compatibility or guaranteed privacy.