Case study · Research
Digital Cryptomnesia in an Artificial Agent
A preprint by Luiz Ribeiro on two times an AI memory system lost track of where its own information came from, and how each was fixed.
The problem
Cryptomnesia is when a memory comes back but feels like a new idea. Can an AI memory system do the same thing, can it be shown from the system's own records, and can it be fixed?
What we built
- A documented episode in PSIM, checked against the system's own stored records
- A second, system-level mix-up of sources, and a fix that can be checked mechanically
- The full paper, with its LaTeX source, under a CC BY 4.0 licence
What it does now
Version 1.0 was released in August 2026 on Zenodo. Version 1.1 is prepared but not yet released, and an arXiv submission is waiting for endorsement. About eleven weeks before the paper, other researchers published related work on memory in AI agents (MemIR, arXiv:2605.25869).
The stack
- Zenodo DOI 10.5281/zenodo.21880453
- Written in Markdown and LaTeX, built with pandoc
The outcome
- A published preprint, sole author Luiz Ribeiro
- Patent pending (US provisional application on file)
Built with the Harmony fleet
Written by Luiz Ribeiro. The fleet checked the paper's links and dates against what each link actually resolves to.