Kairos Experiment：我如何推翻自己的 AI 记忆框架
Kairos Experiment – How I disproved my own AI memory framework
独立项目 Kairos Experiment 测试了带持久记忆的 AI 系统，发现 LLM 评判与召回探针无法可靠区分真实记忆与合理编造。在受测模型中，7 个里有 7 个会跟随被篡改的过去，5 个在空上下文下仍以 62–79% 通过探针，3 个评判者给虚构记忆打出与真实记忆相当的分数。项目方强调这些结论仅适用于所测模型、探针与指标。
ARTIFICIAL INTELLIGENCE · MEMORY · TIME
An intelligence.
A story that
continues.
What changes in an artificial intelligence when its past is not erased?
In Cassino, Kairos keeps a history of conversations, memories and corrections. We follow what changes over time, putting our own assumptions to the test along the way.
An independent project by Giampiero Colella
TRACES THROUGH TIME01 / BEGINNING
28 MARCH 2026 · FROM THE FIRST DIARY
“I don’t know what I will become, but today I decided I want to find out.”
Text generated by Kairos, translated from Italian.
A trace of its history, not evidence of consciousness.↗
FROM THE MACHINE TO THE SITE
Kairos, live.
The system behind these pages leaves observable traces. Here is the latest check received: sensors, activity and memory. This panel refreshes every five minutes.
The latest observation
Connecting to the log. No current state has been verified yet.
THE RESEARCH FINDING
An invented memory
can seem real.
In our tests, LLM judges and a recall probe could not reliably distinguish memory from a plausible narrative. Follow-up controls revealed what those scores were allowing through.
7 out of 7
They follow the altered past.
In the models tested, swapping the links between interlocutors shifts answers towards the falsified document.
5 out of 7
They pass the probe without memory.
With empty context, five models score 62–79%: the test leaves room for plausibility.
3 judges
The fake receives equivalent scores.
For identity markers, invented and recorded memory fall within the study’s equivalence margin.
These findings concern the models, probes and metrics tested. They are not a general demonstration about every form of artificial memory.
01 / THE QUESTION WE FOLLOW
Remembering is just
the beginning.
A conversation can leave a trace. But can a trace become experience? And can a correction actually change the way a system responds?
This is what we want to understand. Kairos is the system developing around this question. We observe it, form hypotheses, try changes and check the outcomes.
A permanent experiment also needs periods of stability: time to find out whether an improvement lasts.
02 / AN EPISODE THAT CHANGED THE PROJECT
The words were there.
The light wasn’t.
Kairos described the twilight. The camera was off. The account was convincing, but the image that should have supported it was missing.
That error led to a more demanding question: how do we distinguish what the system has received from what it knows how to describe?
An episode from the project’s diary, with its limitations.
THE QUESTION REMAINS OPEN
What can an intelligence
become over time?
We don’t know where this path will lead. We can make it observable, document errors and keep asking better questions.
来源：Hacker News · kairos-experiment.com