Mozilla has released an experimental add-on, MemoryCache, featuring a machine learning dialogue system that takes into account the content users access in their browser. Unlike other AI chats, MemoryCache personalizes interactions with users and utilizes important data specific to the user when generating responses. The project's code is distributed under the MPL license. Installation in Firefox is currently only supported manually in 'about:debugging' mode and requires the application of a patch.
All components of MemoryCache run on the user's system without accessing external resources. The knowledge base of MemoryCache utilizes the GPT4All-J model, trained on collections of English language content such as Pile (825 GB of data), ShareGPT (13 MB), and Dolly (660 MB), and is available in versions with 3, 7, and 13 billion parameters. The privateGPT toolkit is used for working with the model, optimized for utilizing large language models on local machines.
The MemoryCache add-on saves viewed pages in the '~\/Downloads\/MemoryCache' directory, from which a periodically running script retrieves them and sends them to the privateGPT handler (ingest.py) to augment the existing language model. This way, the model adapts to the user's interests and can be used for discussing viewed content in text chat with the bot. Users can ask the bot questions about the saved pages in natural language, for instance, to clarify certain facts, and the bot will generate a response using the local model without accessing external services.
Source: opennet.ru
