Auxiliary-predicted Compress Memory Model(ApCM Model): A Neural Memory Storage Model Based on Invertible Compression and Learnable Prediction

TL;DR

The paper tackles the lack of runtime memory in large language models by introducing the Auxiliary-predicted Compress Memory (ApCM) Model, a learnable external memory that decouples storage from reconstruction using an invertible compressor and a learnable predictor. It combines an invertible neural network encoder (IDRP) with a latent split $z = [z_{comp}, z_{aux}]$ and a predictor $\hat{z}_{aux} = g_{\phi}(z_{comp})$ to enable reconstruction from compact storage, while a global memory bank supports content-based read via cosine similarity and frequency-based write. The main contributions are the integrated architecture (IDRP + memory controller), a complete read-write mechanism for dynamic memory management, and empirical evidence showing superior nonlinear reconstruction over PCA on nonlinear data, with discussions on robustness and integration into AI systems. The approach offers a scalable, trainable runtime memory for AI, enabling enhanced long-context processing, personalization, and continual learning in practical deployments.

Abstract

Current large language models (LLMs) generally lack an effective runtime memory mechanism,making it difficult to adapt to dynamic and personalized interaction requirements. To address this issue, this paper proposes a novel neural memory storage architecture--the Auxiliary Prediction Compression Memory Model (ApCM Model).

Figures (7)

• Figure 1: Architecture diagram of the Invertible Dimensionality Reduction and Predictor
• Figure 2: IDRP training loss curves
• Figure 3: IDRP reconstruction examples
• Figure 4: PCA reconstruction examples
• Figure 5: IDRP training loss curves (real data)