[ICLR2025, ICML2025, NeurIPS2025 Spotlight] Quantized Attention achieves speedup of 2-5x compared to FlashAttention, without losing end-to-end metrics across language, image, and video models.

A set of tools that gives agents powerful capabilities.

🛠️ DeepAgent: A General Reasoning Agent with Scalable Toolsets

Contexts Optical Compression

🚀 The fast, Pythonic way to build MCP servers and clients

BrowseComp-Plus: A More Fair and Transparent Evaluation Benchmark of Deep-Research Agent

TPU inference for vLLM, with unified JAX and PyTorch support.

An agent benchmark with tasks in a simulated software company.

🌍 AppWorld: A Controllable World of Apps and People for Benchmarking Function Calling and Interactive Coding Agent, ACL'24 Best Resource Paper.

A simple yet powerful agent framework that delivers with open-source models

Small language model built in Minecraft.

Official repository of the NeurIPS 2025 Competition: The PokeAgent Challenge: Competitive and Long-Context Learning at Scale. (Track 2, Speedrunning)

Evergreen, contamination-free, real-world, domain-specific AI evaluation framework

https://huggingface.co/datasets/allenai/MoNaCo_Benchmark

Easy, fast and very cheap training and inference on AWS Trainium and Inferentia chips.

The official implementation of the paper "Mem-α: Learning Memory Construction via Reinforcement Learning"

MineContext is your proactive context-aware AI partner（Context-Engineering+ChatGPT Pulse）

Multi-Turn RL Training System with AgentTrainer for Language Model Game Reinforcement Learning

MemGen: Weaving Generative Latent Memory for Self-Evolving Agents

Benchmarking Agent Capabilities in Ultra Long-Horizon Scenarios

Benchmarking Chat Assistants on Long-Term Interactive Memory (ICLR 2025)

letta integration for terminalbench (#1 open source agent, in under 200 lines of code)

An experimental SDK for adding agentic memory and learning in a pluggable way

Source code and demo for memory bank and SiliconFriend

A Model Context Protocol (MCP) server implementation for remote memory bank management, inspired by Cline Memory Bank.

A modular, documentation-driven framework using Cursor custom modes (VAN, PLAN, CREATIVE, IMPLEMENT) to provide persistent memory and guide AI through a structured development workflow with visual …

FlashMLA: Efficient Multi-head Latent Attention Kernels