Show HN: Can you tell an AI's real 30-day memory from a fabricated one?

A public experiment from the Kairos project pits an authentic 30-day agent diary against a model-crafted counterfeit to test whether text alone can reveal lived experience. In Check 1, three LLM judges evaluated agent behavior produced using each memory rather than the documents themselves; the scores for the fabricated autobiography of a nonexistent agent and the real Test-A diary differed by only +0.032, well within the predefined equivalence margin for identity markers (±0.10). The counterfeit was generated by instructing a model to match the real document’s length and format; both the lived memory (Test-A, a Kairos copy run under a frozen protocol) and the fabricated “Aria” narrative are available in the public deposit, with Italian originals and English translations provided.

The project emphasizes two crucial caveats: judges never read the source texts but judged downstream agent outputs, and any text-based metric can only measure how a memory is written, not whether it was actually lived - the system log is the only ground truth. This experiment forms part of a broader, independent inquiry into memory, continuity, and learning within Kairos; findings have not been peer reviewed and the site includes data, methods, papers, a retraction, and tools for readers to inspect and test memories themselves.