How does this differ with code-reveiw-graph

[zer09]

I'm curios if how does this differ with code-review-graph? The quality of the result and the token savings? how does this handle code changes? code-review-graph using git to detect code changes and use that to update the graph.

Just curios.

Thank you

[wenyubo2008]

me too

[DeusData]

Hey @zer09, @wenyubo2008 — great question, and a fair one to ask since both are tree-sitter graphs behind MCP. code-review-graph is a solid project, so I'll keep this honest.

The short of it is they're built for different things. code-review-graph is laser-focused on code review — when a file changes it traces the blast radius (callers, dependents, tests) so your agent reads only the review-relevant set. codebase-memory-mcp is a more general structural memory of your codebase: call-path tracing, architecture maps, dead-code detection, cross-service HTTP/gRPC/GraphQL linking and even infra-as-code (Docker/K8s) indexing — with review-style impact analysis as just one of the tools (detect_changes, which maps your uncommitted diff to affected symbols with a risk score).

On your change-handling point: both are git-driven and incremental. They diff with git, hash to find dependents, and re-parse only what moved. We do the same but lean on a background watcher that just keeps the graph fresh automatically, so you rarely think about it.

Where I'd genuinely make the case for us: it's a single static C binary with zero runtime dependencies — no Python, no interpreter, no setup — which is why indexing and queries land in a different speed tier (average repo in milliseconds, the Linux kernel in ~3 minutes, sub-millisecond queries). It covers 155 languages, and on top of tree-sitter it runs a Hybrid LSP layer: clean-room reimplementations of the type resolution behind pyright, tsserver, gopls, intelephense and Roslyn, baked right into the binary. That means it actually resolves generics, inheritance and dotted cross-module calls (user.profile.display_name() -> the real definition three files away) without ever spawning a language server — something AST-plus-heuristics can't fully do.

On the numbers: they're not directly comparable, so take any head-to-head with a grain of salt. code-review-graph publishes a nice reproducible eval (big per-question token reductions, 100% impact recall) and is refreshingly honest about its weak spots. Ours (and the methodology behind them) are written up in our README — 10x fewer tokens and 83% answer quality across 31 repos. Different scenarios, both real.

So honestly: if your main job is AI code review with blast-radius receipts, code-review-graph is purpose-built for that. If you want a fast, dependency-free structural memory across a lot of languages with type-accurate call resolution and cross-service graphing, give us a try — installs in one command and indexes before you've finished your coffee. Plenty of people could happily run both. Hope that helps!