headroomlabs-ai/headroom
Headroom is a context compression layer for AI agents, reducing token usage by 60-95% across tool outputs, logs, RAG chunks, files, and conversation history, while maintaining accuracy.
Awesome AI Agents › End-to-End Optimizers
agentscope-ai/Trinity-RFT
Trinity-RFT, from the AgentScope organisation, is a general-purpose framework for reinforcement fine-tuning of large language models, described in arXiv paper 2505.17826 and distributed on PyPI as trinity-rft under the Apache-2.0 licence. Its design decomposes the fine-tuning loop into three coordinated components. The Explorer generates experience data by letting the model interact with an environment, which is what makes the framework applicable to agents rather than only to static preference datasets. The Trainer consumes that data and updates model weights by minimising the configured losses. The Buffer sits between them as a data pipeline covering the whole lifecycle, handling processing, cleaning, augmentation and human-in-the-loop stages. The project addresses three audiences explicitly. Agent application developers use it to train LLM-powered agents and raise their performance in a specific domain, following a workflow-development tutorial. Reinforcement learning researchers use compact, plug-and-play modules to design, implement and validate new algorithms with non-invasive customisation, so a new method does not require forking the training stack. Data engineers build RFT datasets and the pipelines that feed them, using a documented operator interface. Recent releases add SGLang support, fully asynchronous weight synchronisation and scheduling tuned to reduce pipeline bubbles, improved stability for mixture-of-experts training, a colocate mode for single-GPU setups, trainer-driven weight synchronisation, automatic parallelism suggestions, enhanced vision-language model support and an upgrade to verl 0.8.0, alongside support for the Qwen3.5 model series. The repository doubles as a research vehicle: examples correspond to accepted ICLR and ICML papers including CHORD, BOTS, Learn-to-Ask and work on entropy dynamics in LLM reinforcement learning, and the framework was used to train the small CoPaw-Flash agent model published on ModelScope and Hugging Face. Documentation and tutorials are published as a project site in English and Chinese.
https://github.com/agentscope-ai/Trinity-RFT
Also in End-to-End Optimizers
topoteretes/cognee
Cognee is an open-source project that provides scalable and modular memory solutions for AI agents using ECL pipelines, enabling efficient data ingestion, knowledge graph generation, and query capabilities to enhance AI performance and reduce hallucinations.
letta-ai/letta
Letta is an open-source framework for building stateful AI agents with advanced reasoning, transparent long-term memory, and context management, supporting multiple LLM backends and providing a graphical development environment.
langchain-ai/opengpts
OpenGPTs is an open-source platform that replicates and extends OpenAI's GPTs and Assistants API, offering customizable cognitive architectures, extensive tooling, and support for multiple language models and vector databases.
aiwaves-cn/agents
Agents 2.0 is an open-source framework that enables the training and self-evolution of autonomous language agents through symbolic learning inspired by neural network methodologies.
getzep/zep
Zep is an AI memory foundation that uses a temporal Knowledge Graph to enable continuous learning and personalized experiences for AI agents by efficiently managing and retrieving contextual chat and data artifacts.
algorithmicsuperintelligence/optillm
OptiLLM is an optimizing inference proxy that significantly improves the accuracy and performance of large language models on reasoning tasks without requiring any model training, by applying over 20 advanced optimization techniques during inference.
mll-lab-nu/RAGEN
A reinforcement learning framework for training multi-turn reasoning LLM agents on the StarPO algorithm, shipping ten Gym-compatible environments plus diagnostics for echo-trap and reasoning-collapse failure modes.