COMPLETED Research

MemoryCompiler

Memory compression framework for long conversations based on compiler optimization theory

// DESCRIPTION

MemoryCompiler treats dialogue memory compression as a compiler optimization problem, using techniques from program analysis to compress conversation history.

Core Concept

Just as compilers optimize code through multiple passes, MemoryCompiler optimizes memory through:

  • Memory IR: Intermediate representation for entities, facts, and relations
  • Dead Memory Elimination: Remove irrelevant historical information
  • Fact Folding: Merge redundant facts
  • Temporal Compression: Compress time-related information
  • Importance Pruning: Keep only salient memories

Evaluation

Tested on MSC, LoCoMo, LongBench, and MultiWOZ datasets.

// HIGHLIGHTS

  • Novel compiler-inspired approach to memory compression
  • Multi-pass optimization pipeline
  • Comprehensive evaluation suite
  • v1.0 stable release

TECH_STACK

Python Qwen 2.5 PyTorch HuggingFace Transformers

PROJECT_INFO

started: 2024-10-01
completed: 2025-01-01
status: COMPLETED
type: Research