Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Revolutionizing Long-Term Memory in AI: New Horizons with High-Capacity and High-Speed Storage

Hiroaki Yamanaka, Daisuke Miyashita, Takashi Toi, Asuka Maki, Taiga Ikeda, Jun Deguchi

TL;DR

Problem: The dominant extract-then-store paradigm risks information loss and poor cross-task reuse, hindering long-term memory for AI. Approach: Propose STONE (store then on-demand extract), deeper insight discovery, and experience memory sharing as non-parametric, inter-inference memory strategies. Contributions: Formalization showing $|M_{ ext{STONE}}| \le |M_{ ext{ETS}}|$ and experiments where STONE reduces retrieval budget and memory sharing accelerates cross-agent learning, while deeper insight discovery improves outcomes in probabilistic tasks. Significance: The trio enables scalable, flexible, and private long-term memory platforms for self-evolving AI agents, with remaining challenges in storage, recall, infrastructure, and security.

Abstract

Driven by our mission of "uplifting the world with memory," this paper explores the design concept of "memory" that is essential for achieving artificial superintelligence (ASI). Rather than proposing novel methods, we focus on several alternative approaches whose potential benefits are widely imaginable, yet have remained largely unexplored. The currently dominant paradigm, which can be termed "extract then store," involves extracting information judged to be useful from experiences and saving only the extracted content. However, this approach inherently risks the loss of information, as some valuable knowledge particularly for different tasks may be discarded in the extraction process. In contrast, we emphasize the "store then on-demand extract" approach, which seeks to retain raw experiences and flexibly apply them to various tasks as needed, thus avoiding such information loss. In addition, we highlight two further approaches: discovering deeper insights from large collections of probabilistic experiences, and improving experience collection efficiency by sharing stored experiences. While these approaches seem intuitively effective, our simple experiments demonstrate that this is indeed the case. Finally, we discuss major challenges that have limited investigation into these promising directions and propose research topics to address them.

Revolutionizing Long-Term Memory in AI: New Horizons with High-Capacity and High-Speed Storage

TL;DR

Problem: The dominant extract-then-store paradigm risks information loss and poor cross-task reuse, hindering long-term memory for AI. Approach: Propose STONE (store then on-demand extract), deeper insight discovery, and experience memory sharing as non-parametric, inter-inference memory strategies. Contributions: Formalization showing and experiments where STONE reduces retrieval budget and memory sharing accelerates cross-agent learning, while deeper insight discovery improves outcomes in probabilistic tasks. Significance: The trio enables scalable, flexible, and private long-term memory platforms for self-evolving AI agents, with remaining challenges in storage, recall, infrastructure, and security.

Abstract

Driven by our mission of "uplifting the world with memory," this paper explores the design concept of "memory" that is essential for achieving artificial superintelligence (ASI). Rather than proposing novel methods, we focus on several alternative approaches whose potential benefits are widely imaginable, yet have remained largely unexplored. The currently dominant paradigm, which can be termed "extract then store," involves extracting information judged to be useful from experiences and saving only the extracted content. However, this approach inherently risks the loss of information, as some valuable knowledge particularly for different tasks may be discarded in the extraction process. In contrast, we emphasize the "store then on-demand extract" approach, which seeks to retain raw experiences and flexibly apply them to various tasks as needed, thus avoiding such information loss. In addition, we highlight two further approaches: discovering deeper insights from large collections of probabilistic experiences, and improving experience collection efficiency by sharing stored experiences. While these approaches seem intuitively effective, our simple experiments demonstrate that this is indeed the case. Finally, we discuss major challenges that have limited investigation into these promising directions and propose research topics to address them.
Paper Structure (25 sections, 1 theorem, 14 equations, 7 figures, 2 algorithms)

This paper contains 25 sections, 1 theorem, 14 equations, 7 figures, 2 algorithms.

Table of Contents

  1. Introduction
  2. Related works
  3. Prerequisites
  4. Non-parametric (vs. parametric) memory
  5. Inter-inference (vs. Intra-inference) memory use
  6. Cross-task and multifaceted use of experiences
  7. Insight extraction
  8. Experience collection and sharing
  9. Formalization of the Paradigms
  10. STONE vs. extract then store
  11. Deeper insight discovery vs. simple experience replay
  12. Memory sharing vs. self-collection
  13. Simple Experiments Demonstrating the Superiority of the Highlighted Approaches
  14. STONE
  15. Deeper insight discovery
  16. ...and 10 more sections

Key Result

Theorem 1

Under the requirement that information potentially useful for any future task may not be lost and the assumption that the memory size of extracted information is the same to that of original experience, the STONE paradigm achieves the smallest possible memory size among all extract-then-store paradi

Figures (7)

  • Figure 1: STONE-based memory platform.
  • Figure 2: STONE
  • Figure 3: STONE (store then on-demand extract) vs. extract then store.
  • Figure 4: Settings.
  • Figure 5: Rewards for each step.
  • ...and 2 more figures

Theorems & Definitions (2)

  • Theorem 1
  • proof