Spontaneous Retrieval from Long-Term Memory for a Cognitive Architecture

Justin Li, John E. LairdAAAI 2015 · DOI

Introduction

Traditional cognitive architectures, such as Soar and ACT-R, typically rely on deliberate memory retrieval. In these systems, an agent must explicitly initiate a search using a specific cue. However, this approach faces a "knowledge requirement" paradox: to retrieve relevant information, an agent often needs to already possess significant knowledge about what it is looking for. This paper, presented at the 2015 AAAI Conference on Artificial Intelligence, proposes a shift toward "spontaneous" retrieval mechanisms to overcome these limitations.

The Challenge of Deliberate Retrieval

Deliberate retrieval is effective when an agent knows exactly what to look for. However, the authors identify three critical gaps where this method fails:

  1. Timing (C1): Knowing precisely when a search of long-term memory (LTM) is necessary.
  2. Cues (C2): Identifying the correct search terms when the specific target is unknown.
  3. Relations (C3): Understanding the relationship between a known cue and the desired knowledge.

Spontaneous Retrieval via Spreading Activation

To address these gaps, the researchers implemented a spontaneous retrieval mechanism within the Soar architecture using spreading activation.

Figure 1: A semantic network representation of the word 'nightfall' and its components 'night' and 'fall', illustrating the complex relationships in long-term memory.

The mechanism operates on several principles:

  • Contextual Heuristics: Activation spreads from elements currently in working memory to their neighbors in the LTM graph.
  • Automaticity: When the agent is not performing a deliberate search, the architecture automatically retrieves the most highly activated LTM element into a working memory buffer.
  • Safety Constraints: To ensure stability, activation does not spread in cycles, and spontaneous retrievals do not receive a "base-level" boost, preventing any single memory from dominating the retrieval process indefinitely.

The approach was evaluated using the Missing Link word-puzzle domain, similar to the Remote Associates Test. Agents were tasked with finding a "link" word (e.g., night) that forms a compound word with three given stems (e.g., fall, fort, time). The dataset comprised 550 compound words derived from 195 stems.

The researchers tested two primary knowledge representations to see how structure affects retrieval:

  • Single-Link: The agent knows nightfall is composed of night and fall, but there is no direct link from night to nightfall.
  • Double-Link: The agent has explicit, optimized links between stems and their compounds.
Figure 2: Comparison between single-link and double-link knowledge representations. Spontaneous retrieval is most beneficial in the less-optimized single-link structure.

Key Insights and Results

The most significant findings occurred in the Single-Link representation, which represents a more general-purpose, less-optimized knowledge base. In this setup, a deliberate agent must exhaustively search combinations, whereas a spontaneous agent benefits from activation converging on the solution. For example, activation from fall, fort, and time all spread toward night, causing it to "pop" into working memory automatically—mimicking a human "Aha!" moment.

Performance Metrics: * Efficiency: The spontaneous agent solved puzzles in 24.5 decision cycles, compared to 56.1 cycles for the deliberate agent. * Latency: Real-time performance improved from 2631.6 ms to 350.7 ms. * Robustness: When distractors were introduced, the deliberate agent's processing time increased exponentially, while the spontaneous agent remained efficient, solving puzzles in under a second.

Conclusion and Future Directions

Spontaneous retrieval provides a powerful shortcut for cognitive agents, particularly when knowledge is unorganized. While the computational cost of calculating spreading activation is high—accounting for 95% of processing time—the reduction in decision cycles and improved real-time speed are substantial. Future work will focus on optimizing these calculations and better integrating automatic "thoughts" into higher-level reasoning processes.

Summary updated

BibTeX
@inproceedings{li2015spontaneous,
  author = {Justin Li and John E. Laird},
  title = {{Spontaneous Retrieval from Long-Term Memory for a Cognitive Architecture}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2015},
  doi = {10.1609/aaai.v29i1.9224},
}

Loading related content...

Figure