AI Pulse · Under the Radar

If you’ve got developers using AI coding assistants across a large codebase, you’ve probably noticed the same waste happening over and over. Every new session, the AI agent has to re-learn your codebase from scratch. It reads files, figures out how things connect, builds up context, and then throws all of that away when the session ends. Next session, same thing again. That’s not just slow, it’s burning tokens you’re paying for.

A tool called Graft is trying to fix that. It builds a knowledge graph of your codebase, the kind of map that shows how files, functions, and modules relate to each other, and commits that graph into git alongside your code. So instead of rediscovering everything each time, the agent can just check the map. It plugs into Claude Code, Cursor, Codex, and Gemini, which covers most of the tools teams are actually using right now.

The numbers come from a Show HN thread where the creator benchmarked it over 162 runs and reported a 42% drop in token use, a 46% drop in tool calls, and a 60% drop in latency. Those are the claims as posted, not independently verified, but the scale of the testing at least suggests the creator put real effort into measuring it rather than guessing.

If you’re running a shop with multiple repos and a few developers leaning on AI coding tools daily, this kind of persistent memory layer is worth watching. Token costs on agentic coding add up fast once you scale past one or two people, and anything that cuts repeated rediscovery time is real money back in your pocket. It’s early and unproven at scale, but it points at a sensible direction, giving your tools a memory instead of making them start over every time. This is the kind of thing we build into an AI command centre, where persistent context across tools becomes standard practice rather than a one off plugin.