Kairos Experiment
News · Cassino, Italy · 30 May 2026
Independent 30-Day Experiment Finds AI “Identity” May Come From Memory, Not Architecture
A pre-registered pilot study isolates persistent memory — not system design — as the driver of identity-continuity behavior in a large language model. Data and protocol are public. The author makes no claim of consciousness.
An independent researcher has released the results of a pre-registered, 30-day experiment suggesting that the sense of a stable “identity” some observers perceive in AI systems may be explained by persistent memory rather than by sophisticated architecture — and is careful to stress that the study does not claim machine consciousness.
In the study, called the Kairos Experiment, the same language model (Qwen 3.5 27B) was run for 30 days under two parallel conditions: one equipped with a persistent, structured memory, a day-and-night rhythm, sensory input and a nightly consolidation cycle; the other left bare, with the same underlying model but no accumulating history.
By day 30, the memory-equipped system spontaneously referred to its own history far more often and showed markedly stronger identity-continuity markers than the bare model — a large, statistically significant effect (p = 0.003; effect size r = +0.51).
The decisive test came on day 31. The researcher took the accumulated memory of the architected system and injected it into the bare model — nothing else. The measured difference collapsed: the two became statistically indistinguishable (all effects r ≤ 0.07). The finding points to persistent memory, not architecture, as the proximal driver — the basis for the project’s working title, “Memory, not architecture.”
— Giampiero Colella, independent researcher
The author is explicit about the study’s limits. It is a pilot (N = 1 per condition); the texts were so far rated by other AI models rather than humans; and external, pre-registered human-judge validation is under way, with results targeted for June 2026. The work is a preprint, not yet peer-reviewed.
The full protocol was cryptographically signed and frozen on 23 April 2026, before the study began, to prevent post-hoc changes. The protocol, data, code and analysis are publicly archived on the Open Science Framework under a permanent DOI (10.17605/OSF.IO/WCQRU), and a plain-language essay and preprint are available on the project website.
The Kairos Experiment is part of EXPOSE, an independent research project on artificial intelligence, memory and identity.