vc.ru Author Proposes Storing AI Memory in 'Folders of Human Values'
On October 30, 2026, vc.ru published a text by Tanya Tsykova on how to combat assistants' 'digital amnesia': long-term memory is proposed to be divided into isolated categories.
On October 30, 2026, vc.ru published a text by Tanya Tsykova titled 'Manifesto of the New AI: How to Defeat Assistants' "Digital Amnesia" with Folders of Human Values.' This is not a product announcement or a study, but an architectural proposal. The author describes the current long-term memory (Memory) function in assistants as a 'chaotic dump of facts without structure.' According to her formulation, when the context window limit (the amount of text the model retains in a single dialogue) is exceeded, 'digital amnesia' sets in: the beginning of the conversation is erased, and the user re-enters context about their life.
The difference from the current approach lies in the principle of organizing memory. Currently, according to the author, developers are following the path of scaling up power: more data, more parameters, an 'omniscient superintelligence.' Instead, she proposes designing memory in the image of a human, not 'God.' Specifically, dividing long-term memory into isolated folders of fundamental human values. A background algorithm should automatically tag (label) information and distribute it across five independent micro-databases. The author calls this a 'revolutionary approach,' but the text contains no description of a working prototype or measurements.
In practice, this would mean that important facts are stored forever and used automatically, even if the chat itself is reset. The note includes a quote from a user's perspective: 'As a user, I want my AI assistant to have structured long-term memory divided into key life categories.' The process is described as three automatic steps that require no action from the person. The source does not provide details of these steps: the list breaks off at the phrase 'requiring no.' The specific five categories are also not specified.
It remains unknown how the author proposes to resolve conflicts between folders and who will determine 'fundamental values.' There is no data on tests, timelines, or implementation in the source.