does Serena functionality overlap with code graphs?

[charger89]

Would code graph solutions like codegraph or codebase-memory-mcp provide similar/same features to Serena, or do they complement one-another?

[MLuc24]

The two projects you've named sit differently to Serena, so it's worth splitting them rather than answering as one.

Against a code graph, it's complementary, and the boundary is what the answer is derived from. Serena's symbolic tools are GetSymbolsOverviewTool, FindSymbolTool and FindReferencingSymbolsTool, and every one of them is a passthrough to a real language server — there's even a RestartLanguageServerTool for when one wedges. So "where is this defined" and "what references it" come back exact and current, because the language server has parsed the file as it is on disk right now, with the same understanding your IDE has. The flip side is that it's a per-query, per-symbol view: you ask about one thing and get precise facts about that thing.

A graph is precomputed and aggregate. It's built to answer "what's connected to what across the whole repo" — transitive dependents, blast radius, clusters — which no sequence of find_referencing_symbols calls gives you cheaply. The cost is that it's a snapshot: accurate as of the last index, and approximate where the extractor didn't fully understand the language. So they trade the opposite way, and using both is reasonable — graph for orientation, LSP for anything you're about to act on.

codebase-memory-mcp is a different story, because Serena already has that. memory_tools.py ships WriteMemoryTool and ReadMemoryTool, described as writing "information about this project that can be useful for future tasks" as markdown. That's the same job: durable project notes an agent accumulates and reads back. Running both means two memory stores that don't know about each other, and the agent will write to whichever tool description it finds more compelling in the moment — which is worse than picking one, since neither ends up complete.

So if you're weighing all three, I'd say Serena plus a code graph is a genuine pairing, while Serena plus a separate memory MCP is a duplication to resolve rather than a stack to run. Same conclusion as the language-server question — check whether the thing you're adding is a different layer or the same layer twice.