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

THE PROBLEM

LLMs lack runtime memory for dynamic personalized interaction

HOW IT WORKS

Invertible Dimensionality Reduction and Predictor (IDRP)

DIAGRAM

Memory Read and Write Pipeline in ApCM Model

This diagram shows how Auxiliary-predicted Compress Memory Model (ApCM Model) encodes queries, retrieves from the Memory Bank, and applies the write policy based on access frequency.

DIAGRAM

Training and Evaluation Setup for ApCM Model

This diagram shows how Auxiliary-predicted Compress Memory Model (ApCM Model) is trained on synthetic or real data and evaluated against PCA and Key-Value Memory Network baselines.

PROCESS

How Auxiliary-predicted Compress Memory Model (ApCM Model) Handles a Memory Interaction Session

1. 01

   Invertible Network Encoder

   Auxiliary-predicted Compress Memory Model (ApCM Model) uses the Invertible Network Encoder with stacked affine coupling layers and permutation layers to map input x to latent z. This guarantees exact invertibility so later reconstruction via IDRP remains lossless in the transform space.

2. 02

   Latent Space Decomposition and Prediction

   ApCM Model splits z into zcomp and zaux, then applies the auxiliary predictor gϕ with a Linear SwiGLU Linear MLP to estimate zaux from zcomp. This forces zcomp to retain enough information for accurate auxiliary prediction.

3. 03

   Memory Read Mechanism

   ApCM Model encodes a query into zcomp, computes cosine similarity with each Memory Bank slot, and selects the highest similarity slot. The selected zcomp is then passed through IDRP to reconstruct xmem and update access frequency.

4. 04

   Write Mechanism

   ApCM Model encodes batch inputs into zcomp vectors, averages them into z̄, and writes into the Memory Bank using a first idle then least frequently used policy. The selected slot is overwritten with z̄ and its access count reset.

KEY CONTRIBUTIONS

Key Contributions

• 01

  ApCM Model architecture

  Auxiliary-predicted Compress Memory Model (ApCM Model) introduces the Invertible Dimensionality Reduction and Predictor (IDRP) that splits z into zcomp and zaux and learns gϕ to predict zaux. This optimizable lossy reconstruction paradigm is trained end to end using reconstruction loss.

• 02

  Memory Read Write Controller

  ApCM Model designs a Memory Read-Write Controller with a global Memory Bank, cosine similarity based read mechanism, and access frequency based write policy. This supports content based retrieval and frequency based updates for runtime memory.

• 03

  Comparison with Key Value Memory Network

  ApCM Model compresses storage dimension from 1024 to 128 while achieving test MSE 0.987171 versus 1.001440 and MAE 0.765991 versus 0.798507 on random data. This demonstrates higher storage efficiency with slightly better reconstruction than the Key-Value Memory Network.

RESULTS

By the Numbers

The random data comparison table evaluates Auxiliary-predicted Compress Memory Model (ApCM Model) against a Key-Value Memory Network on reconstruction error, storage size, and inference time. These results show ApCM Model trades 0.1790 seconds extra latency for 8x compression and small error reductions in MSE and MAE.

BENCHMARK

By the Numbers

The random data comparison table evaluates Auxiliary-predicted Compress Memory Model (ApCM Model) against a Key-Value Memory Network on reconstruction error, storage size, and inference time. These results show ApCM Model trades 0.1790 seconds extra latency for 8x compression and small error reductions in MSE and MAE.

BENCHMARK

Comparison between ApCM Model and Key-Value Memory Network

Test MSE Error on random data for compressed versus full memory storage.

KEY INSIGHT

WHY IT MATTERS

What this unlocks for the field