Memory Structure Web: A Structured Memory Architecture for Confusion in Long-Context Memory

Abstract

Starting from the phenomenon of "remembering, but not matching up" in long-context memory, this paper proposes a possibility: the problem of large language models' long-term memory may not necessarily be information forgetting, but may instead occur in the processes of information compression, retrieval, and recombination—that is, the information itself still exists, but the characters, situations, conditions, times, and developmental relationships become confused, causing correct information to be recombined into incorrect facts. To address this problem, this paper proposes an engineering conjecture called the "Memory Structure Web": a relational structure layer is added between automatic summarization and raw memory, used to preserve the characters, situations, conditions, times, sources, and developmental relationships among pieces of information. When invoking memory, the system first locates relationships through the structure, then constrains the retrieval scope, and finally invokes the raw context for restoration. Meanwhile, for high-value long-term memories, a human-machine collaboration mode of "model generates structure → user modifies and confirms → save" may be adopted. The core viewpoint of this paper is: long-term memory needs to preserve not only "what things are," but also "what relationships these pieces of information originally had." Whether the Memory Structure Web can effectively reduce incorrect bindings in long conversations and improve the accuracy of memory invocation still requires further engineering practice and experimental verification.