vesicularia

a short analysis on unforgettable

nforgettable is a pretty straightforward ability on paper: upon activation, the user will remember everything they experience from that point forward. However, the actual effects of such memory related abilities are often understated and are treated merely as a buff to the character’s existing intellect instead of a rewiring of their body’s core mechanisms. In this series of essays, we’ll explore the nature of memory, its relationship to attention, and attempt to construct a more coherent picture of the mechanisms of Unforgettable.

> Spoil(er) warning: this essay will spoil most of Undead Unluck. If you haven’t read it yet, I highly highly recommend reading it.

Undead Unluck is a manga by Yoshifumi Tozuka about killing God. The setting can be summarized as:

  1. God created rules for the universe (UMAs) that define and embody the concepts we see (e.g UMA Clothes represents clothes, if you killed them the concept of clothes would cease to exist. Similar to devils in Chainsaw Man).
  2. Negators negate the aforementioned rules of the universe (e.g Undead (不死) is immortal in every sense of the word, anything that Unbreakable (不壞) makes is unbreakable). The set of negators is injective on the set of UMAs.
  3. Negators are the players in a game to kill God. Until they succeed, the world is on a loop where everyone dies and respawns without any of their memories. The first seat of the Roundtable and the negator Undead are exceptions to this.

In this analysis, we’ll use the feats demonstrated by Nico Vorgeil, the only bearer of Unforgettable shown in the series. Note that his abilities in Loop 100 are a subset of his abilities in Loop 101.

Before we begin, I’d like to establish a few assumptions that we’ll make:

  1. Nico, being a canonical genius, would know everything that we know and will discuss. Given that negator abilities are dependent on the interpretation of the user, this establishes the lower bound of his abilities.
  2. In Loop 101, most of his feats are enabled by “awareness of the soul” (soul hax). While this functionally removes the biological limits of his abilities (e.g. storing memories in the soul rather in the brain) and makes placing an upper bound on storage capacity impossible, we can use those feats to place an upper bound on the purely biological Loop 100 version of Unforgettable.
  3. Even prior to Unforgettable activation Nico was already an insane polymath scientist and engineer, and acquiring the negator ability just made him even more insane. The average person acquiring Unforgettable will probably not reach his level.

Background

Before we can look at how Unforgettable works we’ll establish a baseline understanding of how normal human memory works first, which is usually through extreme filtering lossy compression.
Every second, you take in approximately a billion bits/second of raw input across all your senses (sight, smell, touch, hearing, taste). These inputs are immediately filtered down to 10-12M bits/second via your biological hardware (nerves, retinal circuitry, cochlear processing, etc), of which 10M is from your eyes alone. This is then compressed by attention into approximately ~10 bits/second, that you can consciously process (zheng and meister). Your FOV, while being ~200 degrees, only provides ~2 degrees of high resolution visual input. What your eyes do is basically like how a CRT TV projects onto the screen: it flits the small window of high resolution input - saccades - across the room. A lot of your visual input is filled in by your brain, kinda like how diffusion models fill spaces in with what they think “should” be there
. Your picture of the world is actually composed of trans-saccadic memory, but you only retain like 3-4 items across eye movements.


“Subconsciously, you filter, prune, discard - forget - information you perceive for your own convenience.”1 - Chapter 113, Page 10

So at this point the information that you can consciously process is already massively reduced from the raw input you can obtain from your senses. But the lossy compression doesn’t stop there - bits deemed unimportant by your attention will also be dropped while processing, as demonstrated by the “gorilla in the crowd” experiment.


Gorilla in the crowd experiment - most audience members asked to track “who has the ball” were completely oblivious to the gorilla’s appearance.

During each saccade, input is suppressed; between saccades, you fixate for 200-300ms and most of the stuff you see is discarded unless your “attention” decides to keep it. Try scanning the text on the rest of this page really fast: you can “see” every word in high resolution, but you haven’t “read” it and likely you won’t remember it. That’s the role of attention - it aggressively filters out the amount of information you have to process so your brain doesn’t get overstimulated.

You’ve probably heard of the idea that you can keep 7±2 numbers in working memory, or the 4±1 revision after removing redundancy. You’re also probably aware of brainwaves as a concept. Here, we have to take a little detour exploring the concepts regarding thinking.

So you have a bajillion neurons. To form coherent thoughts you need the related ones to fire together - firing every neuron at the same time (onion) will cause you to have a seizure - but you can’t wire them together physically. That’s physically impossible. So brainwaves and synchronicity makes this work.

So basically, neurons that fire together vibrate together. Since brainwaves fire across your entire brain, every neuron is hit, but only the ones vibrating at that frequency resonate and get “activated”, thus allowing you to form coherent thoughts with different clusters of neurons. At any given time, you have multiple brainwaves vibrating at different frequency bands: delta (<4hz), theta (4-8hz), beta (12-30hz), and gamma (30-100hz). Theta and gamma bands are the ones in brainwaves that your conscious thoughts are composed of, and the 4-7 gamma bursts per cycle neatly translates into the chunks of information you can hold in working memory. The rest do background daemon work, bodily control, and subconscious stuff.

It’s roughly here where attention is placed. Functionally speaking, your brain does a softmax on the stimuli you receive with attentional gain as inverse sampling temperature (more attention -> less random). There are two sources of attentional gain - top down (you’re paying attention to something) and bottom up (monkey see banana, neuron activation), and they kinda just compete for your attention.

Past your attention filter, your hippocampus is responsible for writing information from your short-term memory to your long-term memory. This is a much longer process, and happens in two stages: protein synthesis consolidates the information into meat (local), and system consolidation (global) where it slowly diffuses across your neocortex across days, weeks, decades. At this point, that bit of information is semi-permanently occupying space in your head. The concept of Lebron most definitely occupies physical space in your head. That’s crazy.

Screenshot from 12/12/2024

Consolidation happens during sleep: your hippocampus takes the short term memories you formed during the day and commits it to long term memory during sleep. As you may know, sleep has 4 stages: NREM 1-3 and REM, going light, light, deep, dreaming respectively. Deep sleep is when your memories get committed. NREM is declarative while REM consolidates procedural/emotional/experiences, which is why experiences during your day may appear in scenarios in dreams. Your dreams, however, don’t get consolidated into actual memory because your MCH neurons actively suppress it. memories/experiences are tagged using LC norepinephrine, and LC norepinephrine drops to near zero during sleep. Nice.

On the other side of the store-fetch cycle is retrieval: how a memory goes back into your context window. Given a cue - a picture of some pavement, a distinct flavour, a string of words - your hippocampus retrieves the rest of the pattern from that one fragment, approximating the cue towards the nearest patterns and filling in the missing parts from your priors, then regenerating the activations from scratch. It’s more like noisy activation recomputation than playback/file reading, and it’s stored in a content-addressable way across the different regions of your brain. This is why you can remember things like a day after you want to try to remember it - you missed the cue the first time, but you found the key after.

So what survives all this? The embeddings you pulled from the experience, not the raw footage - plus whatever was surprising, emotional, or high neuron activation in general. You compact the context. And because the recollection process is noisy, it also means that every read is a lossy write: you fill in features that didn’t exist, erase small details which you missed on recall; recalling a memory destabilizes it and your brain rewrites it afterwards errors and all. This is why Nico forgets Ichico’s face over time as well.

So that’s mostly how an experience becomes a memory for a normal person.

Mechanism

Now we can explore the mechanism of Unforgettable. My understanding is that it’s essentially a CPU with a perfected memory system - Unforgettable is what you get when you remove the latency and memory bounds and you’re only left with compute-boundedness.

A few specifics on what I think it does:

> We will further discuss the nuances and capabilities in the third section.

So this itself is quite straightforward. What I want to explore is what it means to think/work/live with Unforgettable, as living with this kind of brain requires you to think with a completely different system.

Take the idea of activation recomputation - it’s a strategy to reduce memory load during LLM training you do because you’re memory bound. Unforgettable inverts this problem - you have one brain to think, but infinite zero-latency storage. So when working with a system like this, obviously you wanna precompute a lot of stuff and change “online compute” to “fetch”, and thus effective throughput multiplies massively for memory-bound tasks and does nothing for compute-bound ones.

As previously mentioned, Nico stated that he doesn’t forget anything AT ALL - every sensation, experience, perception, is written into memory. Specifically, he mentions temperature, humidity, the sight of how Void collapsed, the smell. So pretty straightforward sensation stuff. Usually this information is dropped unless it causes a big neuron activation. Sometimes it happens when you’re doing high bandwidth perception as well, like if one sense was dramatically heightened due to some stimuli, like if you saw a 10/10 baddie at a McDonald’s. One might think that if this was how you’re perceiving the world 24/7, you’ll be stunlocked and paralyzed from overstimulation; however, since recording is automatic and unconscious, it doesn't compete for the ~4-slot attention bottleneck, so storing everything never floods consciousness. Moreover, with Unforgettable your thoughts are all recorded as well, so he probably has a mental layer between “thinking” and “speech” where he fetches from his CoT instead of outputting directly.


Chapter 113

Now we’ll take a tangent by exploring the Shereshevsky problem. This was an autistic Russian man who apparently remembered everything via synesthesia. This was not a good thing to live with - he was crippled by his memory abilities since he couldn’t abstract away anything and duplicated a significant amount of information (every instance of a person’s face was stored as a new file). If we take what he says to be true, then he traded away the ability to abstract for perfect recollection. Of course, this was during the Soviet Era, it can’t be rigorously proven, but all you need to know is that he ended up an alcoholic because he couldn’t deal with his brain.

So infinite memory becomes an issue when you don’t distill the information into something useful. With a normal brain, you constantly consciously and subconsciously work through the information you have(as previously established, we have multiple brainwave frequency ranges). To conceptualize this more clearly, think about some arbitrary person you know really well. Your mom or something. If I told you to respond as your mom would to “a naked man walks into a knife she’s holding” - a scenario she probably has never encountered (hopefully unless your mom is Fuuko Izumo) - you could make a reasonable guess. You probably never consciously tried to model your mom in your head, but this model exists regardless. This is done automatically by your brain either as part of compaction or world modeling. With Unforgettable, this still happens, and he has a set of embeddings as well as the different permutations of the same thing over time. The tag only touches storage, so his abstraction machinery is fully intact - he gets raw perception and gist.

The Astral Dolls he uses are an extension of this concept. He can automatically “fetch” his mental models of the Union members - they’re most definitely extremely accurate given his perfect information and smarts - and reflexively render them in reality. He literally “tells them” (himself) to “capture Undead” and they act accordingly. Zero latency recall means he can do these things by reflex. Similarly, when he fights Andy using Shen’s Shin-Hakkyoku, he also fights by reflex: when you fight, you will have a high level plan but your actions on a move-by-move granularity are highly reflexive. That’s why you drill pads, hit the bags, drill positions; Nico oneshots this with lossless motor-program storage. He can use procedural memory / motor automation, which offloads execution to the basal ganglia and cerebellum (the "outer brain") and drops prefrontal load to near zero. As long as he knows which muscles to contract and how to move (highly likely from close observation and testing), he can replicate the movements by reflex (albeit with less strength).


Chapter 113

That’s about it for the baseline abilities of Unforgettable. In the third section of this essay, we’ll look at the specifics of the ability and explore some edge cases, such as dreams, drugs, and lobotomy.

  1. “お前らは知らず知らず情報を削ぎ落とし忘れ自分の都合のいいように糧にする”↩