MemAgent
Basic Information
This repository implements MemAgent, a research and engineering framework for training and running memory-augmented agents that enable long-context processing for large language models. It provides the MemAgent architecture and a multi-conversation reinforcement learning framework (RLVR and an extension of DAPO) to optimize agent workflows end-to-end without changing underlying model architectures. The codebase includes training scripts for RL-MemAgent models, quickstart utilities for local vLLM deployment and online LLM services, dataset preparation and evaluation pipelines based on HotpotQA and synthetic long-context QA, and engineering components such as sync/async agent interfaces and Ray-based process pools. It also documents reproducibility steps and offers pretrained RL-MemAgent-14B and RL-MemAgent-7B checkpoints. The repo is aimed at researchers and developers who want to build, train, evaluate, and reproduce RL-driven memory agents for arbitrarily long-context tasks.
Links
App Details
-
$$$$ -
$$$$