On Machines That Forget

Abstract

In the third week of June 2026, three unrelated AI systems reached for the same word: engram, the name a forgotten zoologist gave in 1904 to the physical trace a memory leaves behind. None of the three knew the word's biography. It belongs to one of the great failures in the history of science, thirty years of a man cutting into brains and finding nothing, and to the laser, sixty years later, that finally caught what he could not. This essay is about why the machines forget, what it costs, and why the word they borrowed is more honest than the products wearing it. The short version: a language model rebuilds its entire memory of a document from scratch every time you ask it anything, at a scale that beggars belief, and the companies promising to fix this are mostly fixing the half that was never hard.

---

The man who could not find a memory

Karl Lashley spent the better part of thirty years looking for a single memory. He would teach a rat to run a maze, then remove a piece of its brain, then see what the rat had forgotten. The plan was simple and the logic was sound: somewhere in that tissue sat the trace of the maze, and if he cut in the right place, he would carry it out on the end of a scalpel and finally hold a memory in his hand.

He never found it. What he found instead was maddening. It did not matter much which part of the cortex he removed, only how much. The memory was not in a place. It was smeared across the whole, present everywhere and locatable nowhere, and after three decades of the most patient destruction in the history of psychology, Lashley wrote one of the most quietly devastating sentences a scientist has ever published about his own life's work:

"I sometimes feel, in reviewing the evidence on the localization of the memory trace, that the necessary conclusion is that learning just is not possible."

The thing he was hunting had a name. A German zoologist named Richard Semon had coined it in 1904: the engram, the lasting physical mark an experience burns into living tissue. Semon gave it a sister word too, ecphory, for the act of waking that mark back up. Both terms were so far ahead of any instrument that could test them that Semon was ignored in his lifetime and half-forgotten after it, rescued only decades later by psychologists who realised he had seen the whole shape of the problem before anyone had the tools to touch it.

I am telling you about a dead zoologist and a frustrated rat surgeon because in one week of June 2026, three different pieces of AI software grabbed his word, and not one of them seems to know what it costs to keep that promise.

One week, three engrams

On June 23, a startup called Engram walked out of stealth with ninety-eight million dollars, backed by Kleiner Perkins and Sequoia, advised by Andrej Karpathy, already wired into Microsoft and Notion. Its pitch was a single sharp line: the AI you use every day is a "brilliant stranger." It can read anything and remembers nothing.

That same afternoon, Anthropic shipped Claude Tag, an agent that lives inside your Slack as a permanent teammate, and the feature everyone wrote about was its memory, the way it accumulates a understanding of your team instead of resetting to zero every morning. And underneath both, if you went looking, sat a third Engram entirely: an open-source memory server, and beyond that a research paper from DeepSeek using the same word for something built into the guts of the model itself.

Three systems, three layers of the stack, one borrowed name, in one week. The collision is not a coincidence so much as a confession. The whole industry hit the same wall at the same time, and reached for the same century-old word to describe the way out. So it is worth understanding, precisely, what the wall is made of.

Why the machine forgets on purpose

Here is the part that sounds made up.

When a language model reads a document, it builds an internal memory of it called the KV cache, the running notes that let it pay attention to the start of a contract while it writes about the end. That memory is real, and it is enormous. For a model the size of the original GPT-3, holding a single seventy-thousand-word contract in mind costs around four hundred gigabytes of the most expensive memory in the building. Four hundred gigabytes, to think about one document a person could read in an afternoon.

And then, the instant your request finishes, the model throws all of it away.

Not as a bug. By design. The cache is born when you press enter and demolished when the answer lands, so the next time you ask about that same contract, the model reads all seventy thousand words again, from nothing, and rebuilds the same four hundred gigabytes from scratch. Every question. Every time. It is a brilliant lawyer with no long-term memory, re-reading the entire case file each time you walk back into the room, having forgotten that you were ever there before.

That number deserves a moment, because it is genuinely strange. The contract as a plain text file is under half a megabyte. Squeeze it down to the pure information it actually carries, the limit Claude Shannon worked out for English back in 1951, and it is smaller still, around fifty kilobytes. The model's working memory of that document is roughly nine million times larger than the meaning inside it. We have built minds that hold an idea by inflating it to the size of a small data centre and then setting it on fire when they are done.

This is the wall. Not that the AI cannot read the contract, it reads it beautifully, but that it pays the full price of reading it new, on every single query, forever. At the scale of a company running thousands of these a day, the re-reading becomes one of the largest bills in the building. Ninety-eight million dollars of venture money is, at bottom, a bet on making the machine stop forgetting.

The easy half and the hard half

Now we can see what Semon saw, and why it matters for where the money is going.

He split memory into two acts. Storing the trace, and calling the right one back at the right moment. He insisted, against the fashion of his day, that the second act, ecphory, was the hard one. Anyone can pile up traces. Pulling the single relevant memory out of the pile, in the instant you need it, is where remembering actually lives or dies.

A century later, the AI systems split along his exact line, and it is the most useful thing to know when you are deciding whether to believe any of them.

The well-funded startup is working on storage: precompute the durable parts of your company's knowledge once, so the model reloads them instead of rebuilding them from zero. That is a real problem and a tractable one, which is exactly why it can be sold against that nine-million-to-one waste. Whether it works as advertised is unproven, the company has shown a pitch, not a paper, but the target is sound.

The more telling thing is where the working systems spend their effort. The open-source memory server can hold fifty thousand memories without breaking a sweat; storage is trivial. Almost every line of its actual cleverness goes into one question: of these fifty thousand things I remember, which handful belongs in front of the model right now? It scores each memory on how similar it is to your question, how recently it mattered, how often it has proven useful, and it forgets on a curve, a memory a week old counts half as much as a fresh one, the very same shape of forgetting Hermann Ebbinghaus measured in himself in 1885. The engineering is not in the remembering. It is in the choosing.

Claude's memory makes the same confession in a different accent. Strip away the marketing and its memory is almost insultingly simple: plain text files in a folder, which the agent reads and writes like any other file, with one quietly serious feature, every change is saved forever as a version you can audit or roll back. The storage is deliberately dumb. The intelligence is the model deciding which file to open. Once again: the hard part is not holding the memory. It is knowing which one to reach for.

There is a sting hidden in that design, and it is the cleanest proof that these are real memories and not just fast caches. Anthropic warns, in its own documentation, that because the memory persists between sessions, a malicious instruction smuggled into one conversation can be written into the memory and then read back, later, as trusted truth. A cache can never be poisoned, because it does not live long enough. A memory can. The capacity to be betrayed by your own past is, it turns out, the price of having one.

The laser, and the part still missing

Lashley's story does have an ending, and it is a good one. He died in 1958 without his answer, but in 2012 a team at MIT led by Susumu Tonegawa finished the hunt he started. They engineered mice so that the exact neurons firing during a frightening moment would also become sensitive to light. Then they put a mouse in a safe, ordinary box, shone a blue laser into its brain, and watched it freeze in terror at a danger that was not there, summoned out of nothing by switching its own neurons back on. They had found the engram, held it, and flipped it like a light switch. A hundred and eight years after Semon named it.

The AI industry is not at the laser yet. It is at Semon's moment, the naming, the first crude tools, the confident press releases for a thing not yet understood. What these systems do today is closer to a very fast, very well-organised filing cabinet than to the strange, living process real memory turns out to be, in which the act of remembering physically rewrites the memory each time, editing the past a little on every recall. That sounds like a flaw. It is the opposite. A trace that cannot be changed cannot learn. The faint first gestures toward that, a memory that consolidates, decays, reconsolidates, are only now appearing in these products, and they are early.

So when a friend forwards you the breathless thread about AI that finally remembers, here is the honest reading of it. The machine forgets because we built its memory to die at the end of every sentence, which, oddly, is the easy problem, the opposite of the brain's, where memory is so stubbornly physical that a man could spend thirty years and never cut it out. Storing the world is the half you can raise ninety-eight million dollars for. Knowing, in the live instant, which single thing out of everything you have ever known this exact moment calls for, the thing Semon named ecphory and could not build, the thing Lashley went to his grave without, is still waiting for its mouse and its laser.

They borrowed a beautiful word. It carries a warning none of them quoted: that the distance between naming a kind of memory and actually possessing it once ran to a hundred and eight years, and a scalpel, and the patience of a man who, near the end, could be forgiven for wondering whether any of us really remember anything at all.

---

References

1. Semon, Richard. Die Mneme als erhaltendes Prinzip im Wechsel des organischen Geschehens. Wilhelm Engelmann, Leipzig, 1904. Coined engram and ecphory; storage versus retrieval.
2. Lashley, Karl S. "In Search of the Engram." Society for Experimental Biology, Physiological Mechanisms in Animal Behavior (Symposium IV), 454–482. Academic Press, 1950. Source of the quoted verdict.
3. Liu, Xu, Steve Ramirez, et al., and Susumu Tonegawa. "Optogenetic stimulation of a hippocampal engram activates fear memory recall." Nature 484 (2012): 381–385. The 2012 proof.
4. Schacter, Daniel L., James Eric Eich, and Endel Tulving. "Richard Semon's Theory of Memory." Journal of Verbal Learning and Verbal Behavior 17 (1978): 721–743. The rehabilitation of Semon and the storage-versus-ecphory distinction.
5. Ebbinghaus, Hermann. Über das Gedächtnis. 1885. The exponential forgetting curve mirrored in modern recency scoring.
6. Shannon, Claude E. "Prediction and Entropy of Printed English." Bell System Technical Journal 30, no. 1 (1951): 50–64. The ~1.1 bits/character entropy floor behind the ~50 KB figure.
7. Brown, Tom B., et al. "Language Models are Few-Shot Learners." arXiv:2005.14165, 2020. GPT-3's dimensions, from which the ~400 GB cache for a 93k-token document is computed via the standard formula 2 × layers × kv_heads × d_head × bytes.
8. Engram. "Engram Launches with $98M to Build AI That Actually Knows Your Organization." PR Newswire, June 23, 2026; and Kleiner Perkins, "Engram: Giving Enterprise AI a Memory," June 23, 2026. Funding, the "brilliant strangers" framing, the 100 GB figure.
9. Anthropic. "Using agent memory" (Managed Agents, beta), and reporting on Claude Tag in Fortune and VentureBeat, June 23, 2026. The file-based memory store, immutable versions, and the prompt-injection warning.
10. Engram open-source project (github.com/ayvazyan10/engram, v0.1.0) and Cheng, Xin, et al. (DeepSeek-AI), "Conditional Memory via Scalable Lookup," arXiv:2601.07372, 2026. The two further, unrelated systems sharing the name; the scoring pipeline is drawn from the former.

Sources: Nature (2012); Bell System Technical Journal (1951); Journal of Verbal Learning and Verbal Behavior (1978); arXiv (2020, 2026); Anthropic, Kleiner Perkins, PR Newswire, Fortune, VentureBeat (June 2026). The ~400 GB cache figure is derived from GPT-3's published dimensions; all quotations and figures verified via Exa and Tavily on June 24, 2026. Three distinct systems are named "Engram"; only the venture-funded startup is the subject of the funding passage.

Last updated: June 24, 2026