On-Device Distillation for Privacy-Preserving Memory Encoding
Train a distilled small language model on cloud-based memory operation inputs and outputs to enable local, privacy-preserving memory encoding and temporal normalization.
Source paper
Jiaqi Liu, Yaofeng Su, Peng Xia, Siwei Han, Zeyu Zheng, Cihang Xie, Mingyu Ding, Huaxiu Yao
Generated by AI. The paper’s authors did not write these.
Train a distilled small language model on cloud-based memory operation inputs and outputs to enable local, privacy-preserving memory encoding and temporal normalization.
Replace static hyperparameters in memory clustering affinity equations with a lightweight, learnable gating network that predicts topic-aware decay dynamically.
Implement a retrieval-reflection verification loop that dynamically triggers expanded context retrieval passes when a lightweight verifier flags initial context as insufficient.
This paper introduces SimpleMem, a memory framework for LLM agents designed to balance long-term reasoning accuracy with token efficiency. The system replaces raw context retention with a pipeline that actively compresses dialogue into atomic facts, recursively consolidates them into abstract summaries, and dynamically adjusts retrieval scope based on query complexity.
The model uses a three-stage pipeline: (1) Semantic Structured Compression, which uses entropy-aware filtering and coreference/temporal normalization to create standalone memory units; (2) Recursive Memory Consolidation, an asynchronous process that clusters memories based on semantic and temporal affinity to form higher-level abstractions; and (3) Adaptive Query-Aware Retrieval, which classifies query complexity to toggle between retrieving high-level summaries or detailed factual records.