MemAgent: Reshaping Long-Context LLM with Multi-Conv RL-based Memory Agent
This addresses the problem of efficient long-context processing for LLM applications, representing a novel method rather than an incremental improvement.
The paper tackles the challenge of processing infinitely long documents without performance degradation by introducing MemAgent, an agent workflow that reads text in segments and updates memory with an overwrite strategy; it demonstrates extrapolation from 8K training to 3.5M QA tasks with less than 5% performance loss and achieves over 95% on a 512K RULER test.
Despite improvements by length extrapolation, efficient attention and memory modules, handling infinitely long documents with linear complexity without performance degradation during extrapolation remains the ultimate challenge in long-text processing. We directly optimize for long-text tasks in an end-to-end fashion and introduce a novel agent workflow, MemAgent, which reads text in segments and updates the memory using an overwrite strategy. We extend the DAPO algorithm to facilitate training via independent-context multi-conversation generation. MemAgent has demonstrated superb long-context capabilities, being able to extrapolate from an 8K context trained on 32K text to a 3.5M QA task with performance loss < 5% and achieves 95%+ in 512K RULER test.