How much could
a life weigh?

COF gives AI characters a past. The obvious next question is the size of a past: how many autobiographical memories does a person hold — the moments of one's own life, not the knowledge one has accumulated — and how many bytes would they take if each one were written down with its sensory traces? Nobody has measured either. This page says what the research does know, offers a working hypothesis, lets you test your own, and draws the consequence for a file that would hold a whole life.

Orders of magnitude, not measurements. Every figure below says where it comes from.

How many autobiographical memories, according to the research

Francis Galton asked the question first, in 1879, while experimenting on his own associations: how many episodes of my own life can I actually bring back? (Not what he knew — Latin, the names of plants, the streets of London — but what had happened to him.) His answer — the episodes that had escaped oblivion were to be reckoned in hundreds or thousands, certainly not in millions — has held up for a century and a half. The one attempt to measure it directly (Crovitz & Schiffman, 1974; integrated by Crovitz, Schiffman & Apter, 1991) cued people with words, dated each memory, fitted a power law to how memories thin out with age, and integrated it: 224 accessible episodes for the twenty years before the test, a lower bound at the bottom of Galton's range; Dudai, reviewing it, notes that the fitted function most likely underestimates the total. Diaries tell the other side of the story: Willem Wagenaar wrote down about 2 400 events over six years — roughly one a day, by design — and when he tested himself, about a fifth seemed gone; but once given the right cues, he could not say that any event had been completely forgotten. Recognition goes further still: people recognise thousands of pictures they saw once (Standing, 1973). Yadin Dudai's review (1997) puts the ladder in one line: recall of core episodes is counted in thousands, recognition reaches an order of magnitude more, and if you had kept one episode a day you would hold 25 000 by seventy.

So a hundred memories a year is not a measurement, but it is a defensible place to stand: it is the upper end of what Galton counted, a quarter of what Wagenaar found worth writing down, and the order of magnitude Dudai gives for core episodes. Over fifty years it makes about five thousand — the number this page will carry. With cues, multiply by five or ten; with an exceptional memory (the HSAM cases studied by LePort and colleagues recall nearly every day), by far more.

They are not spread evenly over a life

Two things shape where memories sit. Almost nothing survives from before the age of three or four (infantile amnesia). And a disproportionate share comes from the years between ten and thirty — the reminiscence bump, the most replicated finding in the field (Rubin & Schulkind, 1997). After that, routine years leave few distinct memories, until the last few years, which are still fresh. The nine memories that Clara Mercier — the fictional example character shipped with COF, whose file (sixteen memories in all) is the yardstick for every figure on this page — keeps at 65 happen to follow that profile; it is also what a character's presets should follow, and why Clara's childhood preset keeps one memory where her 65-year-old self keeps nine.

Two ways of remembering — and what weighs

Not every autobiographical memory is the same kind of thing. Endel Tulving separated remembering — travelling back, with images, a sound, the feel of a place — from knowing: being able to say that something happened, and tell it, without seeing it again. The first depends above all on visual imagery (Rubin, Schrauf & Greenberg, 2003); people without mental imagery recall their own past with fewer details and less confidence (Monzel et al., 2024), and in the rare condition called SDAM a person knows the facts of their life yet cannot relive any of it (Palombo et al., 2015). Hearing comes second. Smell is special: odour-cued memories are older, rarer, more intense, less often revisited (Willander & Larsson, 2006). Touch has no numbers at all.

This matters for weight, because only remembered memories carry sensory traces, and only sensory traces weigh anything. A known memory is a few lines of text: kilobytes. A remembered one is a picture or three, perhaps a sound, perhaps a few seconds of film: hundreds of kilobytes to a few megabytes. No study measures the share of each; the deduction that most of a life is known rather than relived is reasonable, not proven. The calculator below makes the share a slider.

What about the thousand songs you know?

Music looks like the exception. Most people can replay hundreds of pieces in their head, some of them almost note for note, and a professional musician holds a repertoire that would dwarf the count above. Two things the research says make it fit. First, a known song is not an autobiographical memory: it is knowledge, stored with the vocabulary and the faces one recognises, and it survives when episodic memory goes — the brain regions that hold long-known music are among the last spared in advanced Alzheimer's disease (Jacobsen et al., 2015). The evening you danced to it is the memory; the song is what the memory points at. Second, a song is stored as structure, not as sound: melody, rhythm and words, pinned to a surprisingly exact key and tempo (most people sing familiar recordings at or near the original pitch, within two semitones, and most within 8 % of the original tempo — Levitin, 1994; Levitin & Cook, 1996), with the rest regenerated from a learned grammar of tonality and metre (Huron, 2006; Pearce, 2018). That is why a musician remembers a structured piece far better than the same notes shuffled (Halpern & Bower, 1982), and why a song compresses so much better than an evening: a melody with its lyrics is a few kilobytes of symbols, a megabyte only as a recording. In a file, then, a piece of music is a reference — a title, a few bars, a ten-second trace like the river at night — attached to the memories it belongs to, not a thousand recordings. The envelope below does not move.

The same goes for knowledge in general: COF's memories are autobiographical by design — moments and periods of a character's life — and a character's knowledge of its world lives in its personality and its documents. Nothing forbids an author from writing a piece of knowledge as a memory when the story wants it that way (a secret about its world that only the character holds, learned one night and kept), but that is not what this page counts.

The same shape as the file

Martin Conway's model of autobiographical memory (Conway & Pleydell-Pearce, 2000) describes it as three levels: lifetime periods, general events, and event-specific knowledge — the sensory-perceptual details that make a moment vivid. A COF file holds the same three: periods, moments, and the media attached to a moment. The correspondence was not designed; it fell out of asking what a tool needs. It is reassuring all the same.

• lifetime periods
  the agency years, the years of the AI war — a COF period, with its span
• general events
  the first morning in the studio, the evening on the quay — a COF moment, with its gist, when, where, with whom
• event-specific knowledge
  the light on the desk, the sound of the river — the moment's media: an image, a sound, a clip, a few remembered lines

What a memory weighs, measured

The one set of memories that exists in the format is Clara's: sixteen nodes, thirty-one media, written for the 1.1 example. Rather than guess, here is what they weigh in the file. Her images are mostly 1248 × 832 JPEGs — about one megapixel, the size generators produce today; her one clip is ten seconds at 480p; her one sound is ten seconds.

Measured on spec/examples/10-clara-mercier, COF 1.1. The whole character, with her faces, voice and motion clips, is 21 MB unpacked — 15.5 MB as a .cof.

So in practice a memory with one picture is a quarter of a megabyte; a rich one, three pictures and a clip, is a megabyte. Triple the resolution — a phone photo rather than a generated one, 1080p rather than 480p — and the same memories weigh three times as much. A megabyte per remembered memory is therefore a fair average at today's sizes, with a factor of three either way depending on how much you keep.

Try your own numbers

Every assumption above is a control here. The defaults are the working hypothesis: a hundred memories a year from the age of three, all of them remembered rather than merely known, with a mix of light, standard and rich media sized on Clara's file and scaled to phone-quality media. Change any of it.

What it means for a file

Take the working hypothesis seriously and the envelope of a near-exhaustive character — a life's worth of autobiographical memories, at today's media sizes — is a few gigabytes. That is not a problem. It is a DVD, a few per cent of a phone. The format already handles it: a .cof is a ZIP container (zip64 accepted) whose first two entries identify it and hold the manifest, so a reader knows what is inside before extracting anything; media are extracted on demand; the builder flags a memory with more than three images, and a video over 20 MB has to be acknowledged on its memory. The rest of a character — references, voice samples, a 3D model, motion clips — is far lighter than its past.

A working envelope, not a rule of the format: COF puts no limit on a file. It is what a platform could plan for: a complete character fits on a phone, with room to spare.

An open question: a memory as a latent space

There is a deeper way to ask the question. A human mind does not file its past as five thousand documents. Memories are reconstructed from dispersed traces, each recall colours the next, and the whole is held in connection weights — a latent space, in today's vocabulary — where what resembles what is the structure itself. If one could gather, reconstruct or author a life's worth of memories, the most faithful form for them might not be an archive at all: it might be a model, a set of weights trained on them, in which every link between memories exists without being declared. That is a research programme in itself, and a fascinating one.

It is worth being clear about what it would cost. Weights are opaque: you cannot open them and read a memory. They cannot be edited one memory at a time. And they cannot reliably forget — removing one training example from a model is an unsolved problem (Bourtoule et al., 2021), which means a real person who withdrew consent, or a user who asked for a personal memory to be purged, could not be honoured. Everything COF was built to guarantee — reading, editing, purging, carrying one's own data — lives in the archive.

The brain, as it happens, does not choose. The standard account (McClelland, McNaughton & O'Reilly, 1995) is two complementary systems: a fast episodic store that keeps the day's events separately, and a slow cortical one that integrates them over time into general structure; consolidation is the training. An archive and a latent space, side by side, one feeding the other. COF already carries both kinds of thing for faces — reference images next to the LoRA trained from them — and the same pattern fits memories: the graph as the source of truth, and beside it a consolidation — an adapter trained on the memories — derived, regenerable, discarded and rebuilt when the archive changes, so that a purge is honoured by retraining rather than by hoping.

So here is the study we would like someone to run. Take a few thousand memories — authored for a fictional character, or reconstructed with consent from a real life — train an adapter on them, and measure three things: whether the model recovers the links the graph declares (after, because, about) without being told; whether it finds links the author did not declare and a reader would accept; and whether retraining after a purge leaves no trace of what was removed. A cheaper first step exists today: an embedding per memory, a vector index carried as a derived file — a latent space you can still read back to its source. If the answers are good, the ultimate form of a character's memory may well be a model. It will still need its archive, for the same reason we need ours: to be able to say what happened.

Sources