Breaking01 / AI
MIGRO retains 99.9% of rebuilding quality after model switch

Researchers reported that MIGRO retained 99.9% of full-history rebuilding quality after a model switch when the receiving model continued adding records to inherited memory, according to their published benchmark figures.
The paper presents MIGRO as a long-term memory system that lets successive language models share accumulated experience. Its design separates memory from individual models so a replacement model can read existing records and extend them.
According to the abstract, MIGRO separates a semantic search index, which finds records by meaning, from records that preserve the original evidence. The design preserves observations and their historical relationships exactly. Reversible encoding reduces storage requirements while allowing the original records to be recovered.
The researchers described interfaces that convert selected records into inputs suited to each model's vocabulary and tokenisation, the way text is split into units for processing. The system selects the records to retrieve only after it receives a question.
The paper's model-switch experiment uses memory rebuilt from the full history as its comparison baseline. The researchers compared that approach with continuing to add records to the memory inherited by the receiving model.
The abstract does not name the models used in the model-switch experiment.