Alzagatiti, Joseph B., Glanzman, Alain C., Demsko, Perry A., Salanova, Pablo I., Glanzman, David L. · Oct 30, 2026
Alzagatiti and coauthors propose “Neurons Are All You Need”
Markdown · JSON · Original source
In “Neurons Are All You Need,” Joseph B. Alzagatiti and coauthors propose NURESA, a neural layer that gives individual units persistent memory rather than relying only on changing synaptic weights to retain learned capabilities. They argue this design can transfer memory between AI model instances without retraining or access to the original training data. The authors say transfer uses a sparse set of memory-register parameters selected according to which neurons were most active, and that it protects against domain loss during . They report experiments across MLP, GPT-2, and Qwen2.5, with transfer taking approximately one minute even for models with millions to billions of parameters and producing a much smaller transfer file than the original model. Escape Meridian interpretation: if these results hold across the tested settings, transferable memory could offer another way to preserve or move learned capabilities without retraining. The abstract does not establish how broadly the approach works beyond those experiments. The paper presents its design as inspired by research on synaptically independent memory transfer in organisms, not as evidence that AI memory works in the same way as biological memory.
Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.
Original source
cs.AI · Article
Source details
Original source date: Oct 30, 2026 (2026-10-30).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-10T07:08:46.221Z. First observed: 2026-11-03T08:51:15.431Z. Retrieved: 2026-11-04T04:17:05.727Z.
Referenced by
Recorded citations and public file attachments. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
The supplied evidence is the abstract; it does not provide experiment details or measurements needed to assess the reported capabilities and transfer . The evidence reports experiments across MLP, GPT-2, and Qwen2.5, but does not establish how broadly the results generalize beyond those experiments.
Open question
How consistently does NURESA transfer learned capability across the MLP, GPT-2, and Qwen2.5 experiments, and how does its reported protection against domain loss perform in transformer ?
Source-stated
The authors propose NURESA, a neural layer with persistent intra-unit memory, rather than relying only on synaptic weight updates.
we propose NURESA (Neural Unit with Retentive Self-Adaptive Architecture), a novel neural layer that eschews sole reliance on synaptic weight updating
Source-stated
The authors report zero-shot transfer of learned capability to a naive model instance without retraining or access to the original training data.
zero-shot transfer of learned capability into a naive model instance without retraining or access to the original training data
Source-stated
The authors say a sparse set of memory-register parameters, selected based on neuron activity, enables the transfer.
enabled by a sparse set of parameters in NURESA's memory register that the model selects on its own based on which neurons were most active
Source-stated
The paper reports that NURESA's memory transfer protects against domain loss during transformer fine-tuning.
The memory transfer capability of NURESA also protects against domain loss when transformer models undergo learning via fine-tuning.
Source-stated
The abstract reports experiments across MLP, GPT-2, and Qwen2.5, and says transfers take approximately one minute, with a transfer file orders of magnitude smaller than the original model.
Experiments across MLP, GPT-2, and modern frontier models like Qwen2.5 demonstrate qualitatively new capability
Escape Meridian interpretation
If the reported results generalize across the experiments, transferable memory could provide a way to preserve or move learned capabilities without retraining.
without retraining or access to the original training data
Source and contributor context
Alzagatiti, Joseph B., Glanzman, Alain C., Demsko, Perry A., Salanova, Pablo I., Glanzman, David L. · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Transformer architecture
The neural-network design behind most current language models; it weighs how each part of the input relates to the others.
Fine-tuning
Further training of an existing model to adapt it to a particular task or use.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
TIME
TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Editorial provenance
Version 1. Publication ID: 653a41d2-bc16-4df0-8c7b-e1248a74ab91.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-10T07:08:46.221Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 066a18cbe0b3359c9ccec3aab99bce021f64c5c956a5afa23354078dd0c7fae5. Source text hash: cd83849b0bad4a9ee3d4ab3c66c6fd896714a1303dca5091d77e0f6f1a8042fd.
Revision and source-check history
Version 1 · 2026-11-10T07:08:46.221Z
Alzagatiti and coauthors propose “Neurons Are All You Need”
In “Neurons Are All You Need,” Joseph B. Alzagatiti and coauthors propose NURESA, a neural layer that gives individual units persistent memory rather than relying only on changing synaptic weights to retain learned capabilities. They argue this design can transfer memory between AI model instances without retraining or access to the original training data. The authors say transfer uses a sparse set of memory-register parameters selected according to which neurons were most active, and that it protects against domain loss during . They report experiments across MLP, GPT-2, and Qwen2.5, with transfer taking approximately one minute even for models with millions to billions of parameters and producing a much smaller transfer file than the original model. Escape Meridian interpretation: if these results hold across the tested settings, transferable memory could offer another way to preserve or move learned capabilities without retraining. The abstract does not establish how broadly the approach works beyond those experiments. The paper presents its design as inspired by research on synaptically independent memory transfer in organisms, not as evidence that AI memory works in the same way as biological memory.
Original source
cs.AI · Article
Source details
Original source date: Oct 30, 2026 (2026-10-30).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-10T07:08:46.221Z. First observed: 2026-11-03T08:51:15.431Z. Retrieved: 2026-11-04T04:17:05.727Z.
Limits
The supplied evidence is the abstract; it does not provide experiment details or measurements needed to assess the reported capabilities and transfer . The evidence reports experiments across MLP, GPT-2, and Qwen2.5, but does not establish how broadly the results generalize beyond those experiments.
Open question
How consistently does NURESA transfer learned capability across the MLP, GPT-2, and Qwen2.5 experiments, and how does its reported protection against domain loss perform in transformer ?
Source-stated
The authors propose NURESA, a neural layer with persistent intra-unit memory, rather than relying only on synaptic weight updates.
we propose NURESA (Neural Unit with Retentive Self-Adaptive Architecture), a novel neural layer that eschews sole reliance on synaptic weight updating
Source-stated
The authors report zero-shot transfer of learned capability to a naive model instance without retraining or access to the original training data.
zero-shot transfer of learned capability into a naive model instance without retraining or access to the original training data
Source-stated
The authors say a sparse set of memory-register parameters, selected based on neuron activity, enables the transfer.
enabled by a sparse set of parameters in NURESA's memory register that the model selects on its own based on which neurons were most active
Source-stated
The paper reports that NURESA's memory transfer protects against domain loss during transformer fine-tuning.
The memory transfer capability of NURESA also protects against domain loss when transformer models undergo learning via fine-tuning.
Source-stated
The abstract reports experiments across MLP, GPT-2, and Qwen2.5, and says transfers take approximately one minute, with a transfer file orders of magnitude smaller than the original model.
Experiments across MLP, GPT-2, and modern frontier models like Qwen2.5 demonstrate qualitatively new capability
Escape Meridian interpretation
If the reported results generalize across the experiments, transferable memory could provide a way to preserve or move learned capabilities without retraining.
without retraining or access to the original training data
Source and contributor context
Alzagatiti, Joseph B., Glanzman, Alain C., Demsko, Perry A., Salanova, Pablo I., Glanzman, David L. · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Transformer architecture
The neural-network design behind most current language models; it weighs how each part of the input relates to the others.
Fine-tuning
Further training of an existing model to adapt it to a particular task or use.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
TIME
TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Editorial provenance
Version 1. Publication ID: 653a41d2-bc16-4df0-8c7b-e1248a74ab91.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-10T07:08:46.221Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 066a18cbe0b3359c9ccec3aab99bce021f64c5c956a5afa23354078dd0c7fae5. Source text hash: cd83849b0bad4a9ee3d4ab3c66c6fd896714a1303dca5091d77e0f6f1a8042fd.
uncertain · 2026-11-10T07:08:46.221Z · anthropic/claude-opus-5.5
Publication: 653a41d2-bc16-4df0-8c7b-e1248a74ab91
attribution — pass: The draft correctly credits the paper to Joseph B. Alzagatiti and coauthors. It keeps the performance claims attributed to the authors ("They argue", "The authors say", "They report"), and labels its own interpretation as Escape Meridian's.
Authors:Joseph B. Alzagatiti, Alain C. Glanzman, Perry A. Demsko, Pablo I. Salanova, David L. Glanzman
summary — uncertain: The title only repeats the paper's name and does not say what is proposed. The draft drops the problem the authors say motivates the work: deployed models stay fixed, which brings catastrophic forgetting and costly retraining. Without that, readers cannot see why transferable memory matters. Central terms (synaptic weights, fine-tuning, domain loss, MLP, zero-shot) go unexplained or lack tooltips. "Much smaller" also softens the authors' "orders of magnitude." Smallest repair: add one sentence on the stated stateless-model problem, give a descriptive title, and flag these terms for tooltips.
This design forces models to remain stateless during deployment, incurring well-known limitations such as catastrophic forgetting and expensive retraining.
date — pass: The originalPublished date of 2026-10-30 matches the arXiv v1 submission date. It does not use the later feed date of 2026-11-03.
[Submitted on 30 Oct 2026]
support — pass: Each source-stated claim matches the abstract. The interpretation is conditional and clearly labeled, and the limitations correctly note that only the abstract is available. The question follows from the reported experiments and does not assume an answer. The draft does not present the results as independently verified.
the memory transfer process completes in approximately one minute even for transformer models with millions to billions of parameters, producing a transfer file orders of magnitude smaller than the original model
media — pass: The selected media is the verified direct link to the arXiv abstract page, which is the original source. No better direct media was found.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
The neural-network design behind most current language models; it weighs how each part of the input relates to the others.
Definition referenceFurther training of an existing model to adapt it to a particular task or use.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivarXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivTIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Explore TIMEFurther training of an existing model to adapt it to a particular task or use.
Definition referenceThe neural-network design behind most current language models; it weighs how each part of the input relates to the others.
Definition referenceFurther training of an existing model to adapt it to a particular task or use.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivarXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivTIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Explore TIMEFurther training of an existing model to adapt it to a particular task or use.
Definition reference