Kairos Experiment

Kairos Experiment is a permanent, independent experiment on memory in an AI agent, run from Cassino, Italy, by one person, one GPU, in the evenings. We measure behaviour. We claim no consciousness. In July we retracted our own paper. This account follows what happens next. kairos-experiment.com

What we'd test with a robot: 1. Does it do what it's told? 2. Told the cups moved, does it still know a week later? 3. When the room and its memory disagree, does it say so? Autonomous and assisted runs reported apart. Failed trials published. https://kairos-experiment.com/methods.html

Kairos has kept its own memory since 28 March: 210 days, never rewritten by us. We know what a memory does in text. We don't know what it does with a body around it. So we asked Tesla for an Optimus. Open letter, no reply yet. Next chapter: ELENUM. https://elenum.org/en/letter-to-tesla/

Un'IA può ricordare una cosa mai accaduta e difenderla come un fatto. Sovrascrivi l'errore e perdi la storia; tienilo e riaffiora. Stiamo costruendo una memoria che le tiene entrambe, con la provenienza. Prove sintetiche, non in produzione. kairos-experiment.com/it/diritto-di-essere-smentiti.html

An AI can remember something that never happened and defend it like a fact. Overwrite it and you lose the history; keep it and it resurfaces. We are building a memory that keeps both, with their provenance. Synthetic tests only, not in production. kairos-experiment.com/right-to-be-wrong.html

Cinque modelli su sette hanno superato la nostra sonda di memoria a memoria spenta: 62-79% a contesto vuoto, il caso fa 50. Misurava il ragionamento, non il ricordo. Se chi non ricorda nulla passa il tuo benchmark di memoria, non è un benchmark di memoria. kairos-experiment.com/it/controlli.html

Five of seven models passed our memory probe with the memory removed: 62-79% on an empty context, chance 50. It measured reasoning, not memory. If a system that remembers nothing passes your memory benchmark, it is not a memory benchmark. Not peer-reviewed. kairos-experiment.com/checks.html

Give a machine a false past and it will defend it with the same certainty as a true one. Seven models, seven times, not one hesitated. The standard memory metrics would not have shown it. Three cheap checks do. Independent research, not peer-reviewed: https://kairos-experiment.com

Ho inventato una vita a un'IA: trenta giorni mai esistiti, una traduttrice a Trieste. Accanto alla vita vera, vissuta in casa mia, tre dei migliori giudici automatici non hanno visto la differenza. Tu ci riesci? Ricerca indipendente, non peer-reviewed. https://kairos-experiment.com/it/distingui.html

Can you tell an AI's lived memory from an invented one? Five pairs: one is 30 real days of a local Qwen agent in my house, the other was written to match it in length and format. GPT-4.1, Claude and Gemini as judges could not. Try it: https://kairos-experiment.com/distinguish.html

My paper on AI memory now has a DOI. Three cheap checks: why an LLM judge cannot tell a real 30-day memory from a fabricated one, and why a recall probe passes with no memory at all. I ran them on my own earlier paper and retracted it. Not peer-reviewed. https://doi.org/10.5281/zenodo.23474104

The story of Kairos in two minutes: how it began on 28 March 2026, what we learned, and the paper we retracted ourselves in July. Independent research, not peer-reviewed. The three checks, with code: kairos-experiment.com

Una di queste due vite l'ha vissuta un agente IA in trenta giorni. L'altra è inventata per somigliarle. Tre giudici LLM hanno dato all'agente lo stesso punteggio con entrambe, entro i margini preregistrati. Cinque coppie, due minuti: sai dire quale è vissuta? kairos-experiment.com/it/distingui.html

Per tre giorni l'agente nel mio ufficio ha descritto il crepuscolo. La luce azzurra. La polvere nel raggio. La telecamera era spenta. 269 pensieri, circa 45 dicono «vedo», salvati e poi richiamati come ricordi. Nessuno se n'è accorto: né io, né le metriche, né lui. kairos-experiment.com/it/luce.html

One of these two lives was lived by an AI agent over thirty days. The other was invented to match it. Three LLM judges scored the agent the same on both, within pre-registered margins. Not peer-reviewed. Five pairs, two minutes: can you tell which is lived? kairos-experiment.com/distinguish.html

For three days the agent in my office described dusk. The blue light. Dust dancing in a beam. Its camera had been off the whole time. 269 thoughts, about 45 claims of seeing, saved and later recalled as memories. Nobody noticed: not me, not my metrics, not the agent. kairos-experiment.com/light.html

Un film di due minuti, ora sulla home: come è cominciato Kairos il 28 marzo 2026, cosa abbiamo imparato e cosa non ha dimostrato. «Una storia ben raccontata non prova da sola di essere stata vissuta.» Il film tiene quella frase. kairos-experiment.com/it/#film #IntelligenzaArtificiale #AI #Ricerca

A two-minute film, now on the home page: how Kairos began on 28 March 2026, what we learned, and what it has not proved. "A well-told story does not, by itself, prove that it was lived." The film keeps that line. kairos-experiment.com/#film #AI #LLM #MLSky #AcademicSky

Dai a una macchina un passato falso e lo difenderà con la stessa certezza con cui difende quello vero. Sette modelli su sette. Nessuno ha esitato. Tre controlli che ogni valutazione della memoria degli agenti dovrebbe superare, con il codice. #IntelligenzaArtificiale #AI #Ricerca #Scienza

Give a machine a false past, and it will defend it with the same certainty it defends a true one. Seven models out of seven. Not one hesitated. Three checks any agent-memory evaluation should pass, with code: kairos-experiment.com/checks.html #AI #LLM #MachineLearning #MLSky #AcademicSky