I accidentally hit SOTA on agentic memory by using AI companions (graph.coder.company)

A new tool, graphCTX, has been developed to enhance memory management for AI coding agents, specifically focusing on software development contexts. This innovative local memory solution retains critical repository knowledge—such as commands, conventions, and previous fixes—allowing developers to save time on re-explaining information. By anchoring memory directly to Git commits and avoiding model-based guesses, graphCTX ensures that agents retrieve accurate and relevant coding facts swiftly, with response times averaging under 1 millisecond, compared to nearly 500 milliseconds for established systems like Supermemory. Significantly, graphCTX promotes reliable knowledge usage, extracting facts directly from various repository files, and maintaining them as the codebase evolves. This approach not only minimizes unnecessary noise in memory recall but also enhances developers' trust and auditability of the information used by AI agents. The system's compact and tagged memory structure, along with the straightforward installation process, promises to streamline workflows in coding environments, thus advancing the practical application of AI/ML tools in software development.