The Analysis of AI Memory Mechanism: From SOAR Architecture to ChatGPT Memory Mode

This is a contribution article from Ksenia Se in the Huggingface community, which explores topics related to AI memory, including early cognitive architecture SOAR, types of memory, memory applications in generative agents, memory modes in LLMs, and the potential impact of AI on human memory.

In the field of artificial intelligence, memory plays a crucial role in the Agentic Workflow. Memory is closely related to knowledge and profiling but differs in function and granularity. Profiling defines the identity, behavior patterns, and operating environment of an agent; knowledge provides the basis for decision-making; and memory is a dynamic record of experience that integrates various elements and participates in the decision-making process. Although memory research has gone through decades, the challenge of enabling large language models (LLM) to achieve stable and continuous memory remains.

1. SOAR Architecture: Early Exploration of Agentic Memory Systems

In 1987, Allen Newell, Paul Rosenbloom, and John Laird proposed the SOAR architecture, aiming to build a general artificial intelligence system. The core idea of the SOAR architecture is to integrate problem-solving, learning, and memory into a unified cognitive framework. Its innovation lies in distinguishing between working memory (Working Memory) and long-term procedural memory (Long-Term Procedural Memory). Working memory is responsible for handling immediate…