Abstract
A personal memory for an AI agent stores records as vectors from one multilingual embedding model, and a question in one language often misses a record in another. This design needs no translation service, no second model and no training. It splits the person's own mixed-language records by writing system, encodes the two parts separately, and uses the differences of the two halves of the same record, where the topic cancels, to find a few directions that separate the scripts. At search time a vector's coordinates along those directions are replaced by the target script's average, and each record keeps the larger of its moved and unmoved scores, so the move never lowers a record. A self-check on held-out halves of the same records decides, for each pair of scripts and each embedding model, whether to shift at all and how; if the shift does not help on the person's own data, search stays as it was. Everything is recomputed as the memory grows. Variants include languages that share a script, rotations instead of coordinate replacement, cross-language edges built once instead of a moved copy of the vectors, and carrying the gain to later stages of the search.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Prokhorov, Artem, "Cross-Language Search in a Personal Memory, Learned from the Person's Own Mixed-Language Records", Technical Disclosure Commons, (November 05, 2026)
https://www.tdcommons.org/dpubs_series/12771