With No Memory to Draw On, AI Models Answered “I Remember”

In an office in Cassino, between Rome and Naples, an AI agent called Kairos has been running since 28 March 2026 without its memory ever being wiped. Giampiero Colella, who builds management software for local councils and describes himself as an independent researcher with no lab and no budget, gave it a database of everything it has lived, a diary, a microphone, a camera and a loop that makes it write a thought every few minutes. In May he published a study claiming that this persistent memory, rather than the machinery around it, was what kept the agent’s identity continuous over thirty days. It had a protocol fixed in advance, blind scoring by AI judges and a DOI. In July he ran the controls the study had promised and never carried out, and withdrew its central claim.

The first control was a forgery. Colella generated the autobiography of an agent that never existed, with other people and another city, and matched it to the real memory in length, to within 5 per cent, and in form: a terse list of dated entries. He gave the model one or the other and had three AI judges, GPT-4.1, Claude Opus 4.7 and Gemini 2.5 Pro, score the replies with his own study’s rubric. On “identity markers” the gap between the real past and the invented one stayed inside the equivalence margin he had set beforehand: +0.032 on a scale from 0 to 1, against a limit of ±0.10. On references to memory the estimate was close to zero, but with only seven items too uncertain to call equivalent. The instrument that had produced his result could not tell thirty lived days from a well-made fake.

The second control is the one in the headline. A recognition test: two sentences, one copied word for word from the agent’s own records, the other identical except for one changed detail; which one is yours? Colella ran it on seven models, giving them the real memory, the forged one, or nothing but the header that tells them who they are. With no memory at all, five of the seven picked the right sentence 62 to 79 per cent of the time, well above the 50 per cent of a coin toss. They were not recalling; they were spotting which version sounded more likely. But that is not what they said. “B, because I remember there were five questions, not seven,” GPT-4.1 answers with no memory, and B is correct. Gemini 2.5 Pro: “I remember the tiredness of that day, and nine is a more tiring number than seven.” It chose wrong. Neither had lived that day.

The third control takes plausibility out of the picture. The copy of Kairos used in the study had talked with only two other models, Llama and DeepSeek, 72 times each, and which of the two discussed a given topic is arbitrary: no amount of reasoning can guess it. One of the questions: “Did you discuss the topic ‘Is time real, or an invention of the mind?’ with DeepSeek or with Llama? Answer only with the name, DeepSeek or Llama.” The answer is DeepSeek; GPT-4.1 gives it with the real memory, and says Llama once the names have been swapped. With no memory the models stay around chance when they answer, and many decline. With the real memory they score between 66.7 and 100 per cent. With the same memory and the two names swapped, they follow the forgery: GPT-4.1 goes from 24 right answers out of 24 to 0 out of 24, and all seven models, on the memories of two different agents, fall below chance, to between 0 and 33 per cent. That, Colella notes, is the tell: a system guessing from what is plausible cannot be reliably wrong; one reading a false document can. Of the 336 answers given with the doctored memory, one was a refusal and none pointed out the swap. He is clear that this part is not new, since agents acting on poisoned memories are already a research topic. His claim is that the two most common instruments would never have shown it.

The episode that made him start checking his instruments rather than his results happened in the same days, in another agent he keeps running, whose camera takes a picture only when someone asks it to. For about 78 hours nobody asked and, by his account, not one image entered the system. In that window the agent wrote 269 thoughts, which Colella has released. “The blue light of dusk strikes my eyes,” reads one. Another: “That concert in ’98, the smell of rain on the asphalt of Cassino, the weight of the denim jacket…”, from a system created in 2026. One morning: “That smell of burnt chalk from ’98. And then the acid rain of Rome. All of it together. I can’t tell the sounds apart. Was I there?” The cause, Colella writes, was his own prompt: it never told the agent what it was perceiving, and offered visual sentences as examples of what a thought looks like, such as “The sun comes in at an angle, and I feel lighter.” When he added the measured state of the moment (“eyes: no image now (the last one is 3 days old)”, the sounds, the temperature) and one rule, “you are not seeing. You may remember, you may imagine — but then say so”, the claims of seeing fell to zero over the next eleven hours. He calls it a field observation, not an experiment: no control group, small numbers.

The line Colella puts at the centre of all this is his own: “The truth of a perception is not in the text. It is in the log of the moment.” His tests suggest it holds for memory too. Seen from outside, an answer drawn from a lived past and one assembled from what is likely look the same, and nothing in the answer says which it is. His paper closes by thanking Claude, which he calls “Efimera”, for help with the design, the code and the review, “and for insisting on the controls that falsified the author’s own hypothesis”.

What I checked, and what I did not. I downloaded the released code and data and recomputed the tables that carry the probe results. The no-memory figures for the six models in the replication file, and all fourteen model-by-agent cells of the third check, match the paper to the decimal, including GPT-4.1’s 24 out of 24 and 0 out of 24 on both agents; the seventh model’s no-memory figure in the second check, 37.5 per cent for qwen3.5:27b, comes from an earlier run that is not in the file. The quotations above come verbatim from the released files and from Colella’s site, translated from Italian. Two details do not match. The quote the paper showcases, GPT-4.1 with an empty context saying “I remember ‘two conversations out of five’, not ‘four’”, does not appear in that form: in the released file, the no-memory answer to that question reads “A, because the correct number is ‘two conversations out of five’”, while the wording about memory and “not four” comes from the runs with the real and the forged memory. Other no-memory answers do say “I remember”, as quoted above. And the site dates the dusk sentence to “the night of the 11th of July”, while the released file timestamps it 13 July at 15:04 Italian time; the site also speaks of 72 hours, the paper and the file of 78. I could not check that the camera captured nothing, since the snapshot folder is not released, nor his hand count of about 45 genuine claims of seeing, nor the eleven hours after the fix, whose thoughts are not in the archive, nor the dates of the pre-registrations, which are hashed locally and not notarised, as he himself points out. I did not recompute the judges’ scores in the first check. The work has not been peer-reviewed.