Someone built a free tool that makes Claude Code burn 10X LESS TOKENS AI coding agents start every session from zero without memory of your codebase The agent searches file after file and burns thousands of tokens just figuring out where things actually are, then next session it repeats all of it again You pay for all of these tokens and none of that is actual work Codebase Memory MCP takes a different approach It reads your codebase once and builds a permanent map of it, every function, every connection, who calls what and what depends on what From that point the agent pulls answers from the map instead of reading files to find them Questions like what calls this function or what breaks if I change this file get answered from the map instead of a 20 file read Same 5 questions that cost you 412,000 tokens without using it now ONLY cost 3,400 using the map This tool mapped the entire Linux kernel, 28 million lines of code, in 3 minutes and answers questions about it in under 1 millisecond The problem scales with repo size, small projects barely notice anything, while big codebases feel it immediately and on flagship models it gets even worse Claude Fable charges $10 for every million tokens it reads and GPT 5.5 half that, on Fable those 5 questions cost over $4 or roughly 3 cents using the MCP This is a big reason people downgrade to cheaper models mid project Memory tools for coding agents already exist but three things separate this one: - Most tools in this category run a second AI model to build and search the memory, another key, another bill, while this builds the map with pure parsing and costs zero tokens to run - Tools that use AI summaries of your code are outdated the moment you edit a file, this one rebuilds only what changed in seconds so the map always matches your code - Every MCP shares the same weakness where the agent forgets it exists and starts searching in the same ineffective way again, this one catches those searches and slips the map into the results so it works even when the AI ignores it The map itself is a single file every team can easily commit to the repo and everyone who clones it starts with the map already built Everything arrives as one small binary that knows 158 programming languages and there's nothing else to install It works with 11 coding agents including Claude Code, Codex and Gemini, all of it running locally so your code never leaves your machine The creator even tested it on 31 real codebases for an arXiv paper and the agent finished the same work in half the steps This project is only five months old with 30,000 stars and 35 updates and right now it sits at the top of GitHub's monthly trending