Mnemoria: An inspectable platform for studying how large language models form and revise their picture of individuals
Conversational assistants increasingly keep a memory of the people who use them. They store details from earlier conversations and use them to shape later replies. The person being remembered usually sees none of this. They cannot tell what has been concluded about them, which conclusions reached the model before a given reply, or why a conclusion changed. Mnemoria is a free platform that makes this process visible. While a person talks with a language model, it records what was concluded from each message and whether the person said it or the system inferred it, how those memories are stored, revised and allowed to fade, and the exact context the model received before each reply. It also provides guided experiments, short lessons and a research Lab for controlled runs with exportable traces. Every result is tied to a versioned configuration. The report describes the design principles, the memory pipeline and the evaluations behind its main decisions. In a comparison of two models for memory revision, the smaller model classified only 1 of 9 held contradictions correctly and treated most of the others as changes of view, so revision stayed on the larger model. In a pilot of leave-one-out attribution, removing a single memory usually changed a reply less than running the same model twice on the same input. All evaluations are small, several were one- off development checks whose materials were not archived, and their limits are stated with them.
Authors
- Tunjay Seyidali (ORCID: https://orcid.org/0009-0007-5836-0278)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-17
- DOI
- https://doi.org/10.5281/zenodo.24325625
- Primary Topic
- Explainable Artificial Intelligence (XAI)
- Type
- article
- Field-Weighted Citation Impact
- 0.00