Show HN: Omneum – Local, text-blind context control plane for AI agents (github.com)

Omneum, a new open-source context control plane for AI agents, has been introduced to improve the management of evidence in large language models (LLMs). It addresses the issue of "context rot," where the excessive accumulation of context leads to degraded accuracy in reasoning by using a provenance-aware graph inference system. This approach allows for the evaluation of source reliability and claim support, utilizing dependency tracking to distinguish between genuinely independent information and duplicated claims from multiple sources. Significantly, Omneum enables AI agents to reconstruct and optimize context windows efficiently, selecting relevant passages while adhering to token budgets. For instance, it demonstrated superior claim coverage with reduced context capacity—covering all four claims in two chunks while utilizing only 51 out of 80 tokens. This capability is crucial for enhancing the efficacy and accuracy of AI-generated responses, as it refines how context is curated before reaching the model, ultimately leading to better-informed and more reliable outputs. With its Python SDK and straightforward installation process, Omneum is poised to be a valuable tool for developers and researchers in the AI/ML community.