The World Federation of Neurology has declared July 22, 2026 as the World Brain Day for this year.
The Federation of European Neuroscience Societies, [FENS] held their forum from July 6 – 10, 2026, in Barcelona, Spain. It was hosted in collaboration with the Spanish Society of Neuroscience.
As the race towards advancing artificial intelligence, particularly large language models [LLMs], plow on, there is a sure sign to know when the research labs are approaching artificial general intelligence [AGI] or artificial superintelligence [ASI], it is when a digital memory is developed that can mirror human memory.
Simply, all the AI labs that are looking for another breakthrough that is comparable to the transformer architecture will only achieve something close to what is currently possible, so long the datasets are still the same. It is unlikely to accomplish generalized intelligence with disparate memory. The suggestions of world models, spatial intelligence, neurosymbolic AI, neuromorphic AI and so on, would return niche answers, without a new memory architecture.
The human memory is optimized for intelligence, not particularly for storage. The human memory is mostly for usage, so even when storage, recollection or whatever is possible, what the memory is fully applicable to is usage.
This is different from digital memory, which, from the early eras of digital texts, images, audios and videos, the goal was accuracy or basic storage. It was not expected to be used, beyond just being memory.
Then, as deep learning matured, and it was necessary to train models, digital memory as was available, was used. Then as architectures got better, AI outcomes improved, yet was niche.
Then, even as the excellent transformer architecture blossomed, it gave machine incredible capabilities. Then, as optimization algorithms also grew, it did the same. This is part of the recent advances of the frontier AI models from Anthropic, OpenAI, Google, Meta, the Chinese and so on.
So, at this time, the best of compute has been applied to training AI, the best possible algorithm, yet no AGI, this means that the last sector to solve the problem is memory.
Now, if there is new memory research by any organization that will mirror how human memory is stored, it could become the direction towards how artificial general intelligence [and safety for it] might be achieved.
Human Memory
It is theorized that human memory is stored as collections of information, for the most part. This means that instead of storing unique information about things, the memory seeks what is common between two or more information then adds them into a collection.
For example, table is a collective store. This is where the anything that is similar between all tables is stored. The store can render a table, as is, so that when there is need to use it to think, or discuss and so forth, the regular parts are known or used.
The same for a truck, fan, a chair, an air conditioner, and so forth. Whatever is common is the store, so that it can be used for intelligence. Aside from what is common, there are also overlays between collective sets.
For example, the collective store of table can overlay with that of wood at times, or of a chair, or of a book, and so on. This means that even as some stores cannot collect certain information they still overlay with others where there are similarities.
Overlays are helpful to prepare for intelligence usages as well. Overlays are never permanent. Overlays is the term but it could be a relay alliance, or close. For example, the store of a door can overlay with that of a window.
It is what makes it possible to think of similar things, or be discussing things and say what was not planned but that seems intelligent.
So, the human memory just ensures to prepare itself for excellent use. Assuming every single thing is stored separately, forgetting for humans would be too defective, since there has to be visits to unique things, all the time.
Even with collective storages, there is forgetting, let alone without it. The usage of memory is relays across it, for the foremost objective of intelligence.
This makes it easy to have generalized intelligence for humans. It makes it easy to encounter novel scenarios in reality, and collective storages [or overlays] would be able to ensure interpretations. For example, driving never meets the same situation often.
This means that collective storages and overlays are useful to make instant decisions. It is different from autonomous vehicles, that have to be trained on specific data, yet be unable to generalize situations.
AGI
To build AGI, there would need for new memory architecture for digital data. This would mean that instead of having to store data separately, there would be collective stores then overlays. This would ensure that storage is prepared for training. And it would be possible to generalize data, so that it is not just to show what is available, but to prospect intelligence.
There will be two approaches to improving memory: math and hardware. Math approaches would be to organize stored memory into collections, using Gaussian elimination, eigenvalue decomposition, duality principle, pigeonhole principle and so forth.
At the hardware level, it is possible to explore collective magnetic directions or collective electrical charges of memory cells. It is also possible to have a new VRAM equivalent for storage layers with collections. The objective is to just have data stored like human memory, and then train AI models with the collective data, and their overlays.
Once this is possible, there will be a chance to ensure AI safety and alignment as well, since there will be a collective store of caution and consequences, just like for humans, to make the model safer.
It is possible to build a first architecture for this new model before September 30, 2026, using the postulate in Conceptual Biomarkers and Theoretical Biological Factors for Psychiatric and Intelligence Nosology.
—
iStock image
