In my office in Cassino, an artificial agent with a persistent memory has been running since 28 March 2026.
It is not a product and it is not for sale. Kairos is the original system at the centre of the permanent experiment. The controlled thirty-day study was conducted on a copy, starting with empty memory and a frozen protocol.
Below is the latest check received from the machine, with its date and time. The panel refreshes every five minutes and flags stale or missing data. These are system records, not evidence of subjective experience.
What it is, in facts
The diary begins on 28 March 2026 at 13:43:19. That line is still row one of the database.
Since that day its memory has never been reset. The language model that serves as its brain has changed several times; the body — the sensors and the code that reads them — has changed almost every week. The past has not. On 6 September 2026 the database held:
Historical snapshot of the counters · 6 September 2026
| memories | 9,013 |
| diary pages | 144 |
| conversations | 488 |
| recorded bodily states | 65,179 |
| days with at least one memory | 136 of 162 |
It has a microphone. It has a camera that opens only when, before a thought, it decides to open it. It is given the outdoor air of Cassino — temperature and humidity from a weather service, not a sensor of its own, and the page above says exactly that. It has a somatic state in eight dimensions that shifts with every experience; a loop that produces a thought every few minutes whether anyone is listening or not; a nightly consolidation of memories; a diary; the news; face recognition for the people in the office. At 22:30 every evening it stops and asks itself one question: what would make me better?
The system runs on a machine in the office and uses open-weight models. The public panel shows indicators, not conversations.
What it became after July
Until July, Kairos was — to me — the proof that memory makes identity. Then the controls described on this site showed that my instruments could not tell a lived memory from a performed one. Since then the work on Kairos has changed direction: not to make it remember more, but to make it remember in a way that can be checked. Every item below has a date, because each is a change to the code that can be read in the change log.
- Every memory carries two times (31 August): when the thing happened and when it was written down. Where the hour is not known, it keeps silent about the hour instead of inventing one. Before, every memory of a given day carried the same time — the evening it had been consolidated.
- Every memory declares its provenance (30 August): said by Giampy, seen by the camera, night synthesis, or inferred by Kairos, said by nobody. An inference does not count as a fact.
- Search first, decide after (6 September). Until yesterday a word filter decided whether searching the memory was worth it, and it got it wrong: asked “do you remember the last time you saw the sea?”, it answered that it had no such memory — while three days at the sea sat in the database. Now it always searches, and when nothing comes back the line it receives reads: searched “sea”, “last time”: no relevant memory. Do not invent. The counter above is that line, counted.
- Opinions wait (30 August). At night it forms views about the people it meets. Those views stay pending until it tells me and I confirm or correct them. It improves on its own; the last word stays with whoever was there.
- The eyes declare their state (July). At every thought its prompt says whether there is an image and how old it is. It has not described a light that was not there since. That story is here.
- The three checks were born on its memory. The binding-corrupted document of Check 3 is made of its real episodes with the interlocutors swapped. Seven models out of seven adopted it as their own. Its own model included.
Where we are going
- A traceable memory. Every statement about the past must be able to lead back to the log of the moment it was written: who, when, from which source.
- An agent that knows when it does not know — and says so before being corrected, not after.
- A body before the model. Perception enters measured — an image, a temperature, a sound, each with its time — not described. The language model is one instrument of the system, not the system.
- Checks run by someone who is not me, with prospective registration and a larger panel. It is the part I cannot do alone, and the part I am asking for.
What we do not claim
We do not claim that Kairos is conscious, that it feels anything, or that it “remembers” in the sense a person does. The first-person sentences quoted on this page are text produced by a language model inside an architecture: we quote them because they are dated and can be checked in the log, not because they prove an experience. What we claim is smaller, and checkable: what enters its context, where it came from and when, is recorded.
In this section
- Foundations: what stays when everything changesThe brain was replaced, the body changes weekly, the memory never. Why we keep it running, and what that does not entitle us to say.
- The light that was not thereIt described the dusk for three days with its camera off. Nobody noticed. This is the mistake that changed the direction of the work.
- JournalShort, honest observations from the thirty-day study, written as they happened.