# Revolutionizing Long-Term Memory in Ai: New Horizons With High-Capacity and High-Speed Storage Source: https://www.youtube.com/watch?v=darMYHJIov0 Recap page: https://rapidrecap.app/video/darMYHJIov0 Generated: 2026-02-23T12:03:56.864+00:00 --- ## Quick Overview Kioxia researchers propose the Stone paradigm, which shifts AI memory management from an extract-then-store approach to storing all raw data, enabling far greater long-term memory retention and deeper insight discovery than current methods. **Key Points:** - The Stone paradigm replaces the extract-then-store method by storing all raw data (inputs, outputs, context, noise) in a vector database. - This method allows AI to maintain a comprehensive memory state, avoiding the data loss inherent in current summarization techniques. - The researchers demonstrated that Agent A, using the Stone approach, achieved a 62% success rate on a complex task, whereas the standard agent failed. - The standard approach (Agent A) required 500 trials per agent to learn a task, while Agent B (using Stone) learned exponentially faster. - The primary bottleneck in current AI development is identified as memory capacity and the inefficiency of current retrieval methods. - Stone requires massive, high-speed storage (245 Terabytes for a single drive in their experiment) and avoids the sequential processing bottleneck of current systems. ![Screenshot at 00:00: The opening screen features an illustration of two podcasters with the text overlay "BECOME A MEMBER TODAY!" against a background showing an audio waveform, signaling the start of a discussion or podcast segment.](https://ss.rapidrecap.app/screens/darMYHJIov0/00-00-00.jpg) **Context:** The video discusses a new research paper from Kioxia Corporation by Yumanaka Miashita and colleagues, who propose a radical architectural shift in how AI systems handle long-term memory. This new method, called Stone, aims to overcome the limitations of current extract-then-store paradigms, which are criticized for causing data loss and leading to inefficient learning curves, especially when dealing with complex tasks. ## Detailed Analysis Kioxia researchers propose the Stone paradigm as a revolutionary approach to long-term memory in AI, fundamentally changing how information is retained and retrieved. Instead of the standard method where an AI extracts a summary or snippet of information and stores only that summary (which leads to data loss), Stone advocates for storing the *entire* raw experience—inputs, outputs, context, noise, and even irrelevant data points—in a large, high-speed vector database. This approach is likened to a student keeping every note from every class, rather than just a summary for the final exam. The researchers demonstrated Stone's effectiveness by comparing two agents on a complex task involving navigating a rat maze without food as a reward (pure exploration). The standard agent, which used current extract-then-store methods, failed to learn efficiently, requiring 500 trials per agent and ultimately forgetting crucial details about dead ends. In contrast, the Stone agent, which retained all raw data, learned exponentially faster, achieving a 62% success rate after only 50 trials, validating the idea that comprehensive memory retention is key to deeper insight discovery. The paper argues that the current vector search methods are flawed because they rely on semantic closeness rather than logical context, leading to inaccurate retrievals. By storing everything, Stone allows the AI to perform a comprehensive recall, ensuring that even seemingly irrelevant data points that might relate to a future, unforeseen task are preserved. The main practical limitation identified is the massive storage requirement: the paper suggests 245 Terabytes on a single drive for one agent's memory pool. The authors argue that this shift from compression/summarization to comprehensive storage is necessary to move toward true artificial superintelligence and solve complex engineering problems like designing medical protocols or navigating complex environments. ### The Stone Paradigm - Store all raw data (inputs, outputs, context, noise) in a vector database - Eliminate data loss from summarization - Enable comprehensive recall of every experience ### Experimental Results - Agent A (Standard) failed on the maze task; Agent B (Stone) achieved 62% success after 50 trials, showing 10x faster learning than the standard agent ### Critique of Current AI Memory - Current extract-then-store methods are inefficient, cause data loss, and rely on fuzzy semantic retrieval rather than logical context ### Hardware Implications - The Stone approach necessitates massive, high-speed storage, exemplified by the 245TB requirement for a single agent's memory drive ### Strategic Advantage - Retaining raw data allows AI to connect seemingly irrelevant details later, enabling solutions for future, as-yet-uninvented problems ![Screenshot at 00:00: The opening screen features an illustration of two podcasters with the text overlay "BECOME A MEMBER TODAY!" against a background showing an audio waveform, signaling the start of a discussion or podcast segment.](https://ss.rapidrecap.app/screens/darMYHJIov0/00-00-00.jpg) ![Screenshot at 00:14: The speakers discuss the context window as a bottleneck, mentioning that models cannot process millions of tokens in a single prompt.](https://ss.rapidrecap.app/screens/darMYHJIov0/00-00-14.jpg) ![Screenshot at 00:58: The speakers transition to discussing the research paper, with the title slide implicitly being referenced.](https://ss.rapidrecap.app/screens/darMYHJIov0/00-00-58.jpg) ![Screenshot at 03:25: A visual representation of the Stone paradigm's acronym \(S.T.O.N.E. for Store Then On-Demand Extract\), highlighting the proposed change in retrieval methodology.](https://ss.rapidrecap.app/screens/darMYHJIov0/00-03-25.jpg) ![Screenshot at 08:27: The speaker references the experimental setup, contrasting the standard agent's reliance on policy with Stone's data-driven approach.](https://ss.rapidrecap.app/screens/darMYHJIov0/00-08-27.jpg)