A 120x Token Reduction? How Codebase-Memory-MCP Is Rewriting the Economics of AI Coding

Codebase-Memory-MCP, trending on GitHub, converts an entire codebase into a knowledge graph via a single binary. Its published evaluation shows roughly a 10x reduction in token usage and a 2.1x reduction in tool calls. The widely-cited '120x' figure from the README comes from a narrower micro-benchmark on five structural queries, where about 3,400 tokens replaced roughly 412,000 -- an impressive result, but not something that generalizes to every task.

The similar project CodeGraph leads on GitHub stars (57.3k vs. 25.5k), yet attention has been shifting rapidly toward Codebase-Memory-MCP. Both parse ASTs via Tree-sitter, store results in SQLite, and expose an MCP interface -- but they've taken different paths, one prioritizing performance and deployment form factor, the other prioritizing agent workflows and ecosystem expansion.

Tools like Cursor or Claude Code today understand a codebase by repeatedly grepping and reading files -- a blind-groping process where finding one function call can mean jumping through a dozen files and burning hundreds of thousands of tokens. CodeGraph takes a practical, engineering-heavy route -- solid parsing rules and framework adaptation (route recognition, dynamic dispatch, and so on) -- but maintenance costs climb every time a new framework or complex language semantic gets added.

Codebase-Memory-MCP instead implements lightweight symbol and type resolution in C itself, with no external Language Server dependency, covering call, import, and inheritance relationships across Python, TypeScript, Go, Java, and more. That means when asked 'who calls this interface,' it can return symbol-level, file-level, line-level answers -- higher information density, less noise, and a smaller context for the AI to process.

CodeGraph treats its knowledge graph as a local cache, requiring each developer to re-run indexing locally. Codebase-Memory-MCP instead treats the graph as a code asset: it supports a zstd-compressed snapshot (graph.db.zst, roughly 8-13x compression) that can be committed straight into the Git repo. A new hire pulling the repo gets instant access to full project memory just by loading the snapshot -- no hours-long re-indexing.

With large-model inference cost already the bottleneck, the surrounding infrastructure has to become radically lighter and more efficient -- expect more of it rewritten in C/C++ or Rust going forward. The prediction: future repositories will routinely carry a standardized knowledge-graph file, much like package.json or requirements.txt today.